agora inbox for [email protected]  
help / color / mirror / Atom feed
Re: Problem while setting the fpw with SIGHUP
292+ messages / 6 participants
[nested] [flat]

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-04 08:26  Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-04 08:26 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; pgsql-hackers

At Sat, 31 Mar 2018 17:43:58 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1JSHSPhkHZXj5q2yf3x1MgBN0oYHb9JvcoVh9ZYqB5g+w@mail.gmail.com>
> On Wed, Mar 28, 2018 at 12:10 PM, Kyotaro HORIGUCHI
> <[email protected]> wrote:
> > At Tue, 27 Mar 2018 22:02:26 +0900, Michael Paquier <[email protected]> wrote in <[email protected]>
> >> On Tue, Mar 27, 2018 at 09:01:20PM +0900, Kyotaro HORIGUCHI wrote:
> >> > The current UpdateFullPageWrites is safe on standby and promotion
> >> > so what we should consider is only the non-standby case. I think
> >> > what we should do is just calling RecoveryInProgress() at the
> >> > beginning of CheckPointerMain, which is just the same thing with
> >> > InitPostgres, but before setting up signal handler to avoid
> >> > processing SIGHUP before being ready to insert xlog.
> >>
> >> Your proposal does not fix the issue for a checkpointer process started
> >> on a standby.  After a promotion, if SIGHUP is issued with a change in
> >> full_page_writes, then the initialization of InitXLogInsert() would
> >> happen again in the critical section of UpdateFullPageWrites().  The
> >> window is rather small for normal promotions as the startup process
> >> requests a checkpoint which would do the initialization, and much larger
> >> for fallback_promote where the startup process is in charge of doing the
> >> end-of-recovery checkpoint.
> >
> > Yeah. I realized that after sending the mail.
> >
> > I looked closer and found several problems there.
> >
> > - On standby, StartupXLOG calls UpdateFullPageWrites and
> >   checkpointer can call the same function simultaneously, but it
> >   doesn't assume concurrent call.
> >
> > - StartupXLOG can make a concurrent write to
> >   Insert->fullPageWrite so it needs to be locked.
> >
> > - At the time of the very end of recovery, the startup process
> >   ignores possible change of full_page_writes GUC. It sticks with
> >   the startup value. It leads to loss of
> >   XLOG_CHANGE_FPW. (StartXLOG is not considering GUC changes by
> >   reload)
> >
> 
> Won't this be covered by checkpointer process?  Basically, the next
> time checkpointer sees that it has received the sighup, it will call
> UpdateFullPageWrites which will log the record if required.

Right. Checkpointer is doing the right thing, but without
writing XLOG_FPW_CHANGE records during recovery.

The problem is in StartupXLOG. It doesn't read shared FPW flag
during recovery and updates local flag according to WAL
records. Then it tries to issue XLOG_FPW_CHANGE if its local
status and shared flag are different. This is correct.

But after that, checkpointer still cannot write XLOG (before
SharedRecovoeryInProgress becomes false) but checkpointer can
change shared fullPagesWrites without writing WAL and the WAL
record is eventually lost.

> In general, I was wondering why in the first place this variable
> (full_page_writes) is a SIGHUP variable?  I think if the user tries to
> switch it to 'on' from 'off', it won't guarantee the recovery from
> torn pages.  Yeah, one can turn it to 'off' from 'on' without any
> problem, but as the reverse doesn't guarantee anything, it can confuse
> users. What do you think?

I tend to agree with you. It works as expected after the next
checkpoint. So define the variable as "it can be changed any time
but has an effect at the next checkpoint time", then remove
XLOG_FPW_CHANGE record. And that eliminates the problem of
concurrent calls since the checkpointer becomes the only modifier
of the variable. And the problematic function
UpdateFullPageWrites also no longer needs to write a WAL
record. The information is conveyed only by checkpoint records.

regards,

-- 
Kyotaro Horiguchi
NTT Open Source Software Center





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-06 08:20  Kyotaro HORIGUCHI <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-06 08:20 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; pgsql-hackers

Hello.

At Wed, 04 Apr 2018 17:26:46 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
> > In general, I was wondering why in the first place this variable
> > (full_page_writes) is a SIGHUP variable?  I think if the user tries to
> > switch it to 'on' from 'off', it won't guarantee the recovery from
> > torn pages.  Yeah, one can turn it to 'off' from 'on' without any
> > problem, but as the reverse doesn't guarantee anything, it can confuse
> > users. What do you think?
> 
> I tend to agree with you. It works as expected after the next
> checkpoint. So define the variable as "it can be changed any time
> but has an effect at the next checkpoint time", then remove
> XLOG_FPW_CHANGE record. And that eliminates the problem of
> concurrent calls since the checkpointer becomes the only modifier
> of the variable. And the problematic function
> UpdateFullPageWrites also no longer needs to write a WAL
> record. The information is conveyed only by checkpoint records.

I noticed that XLOG_FPW_CHANGE(fpw=false) is still required by
pg_start/stop_backup to know FPW's turning-off without waiting
for the next checkpoint record. But XLOG_FPW_CHANGE(true) is not
required since no one uses the information. It seems even harmful
when it is written at the incorrect place.

In the attached patch, shared fullPageWrites is updated only at
REDO point and XLOG_FPW_CHANGE is written only for FPW's turning
off before REDO point.

Disabling FPW at any time makes sense when we need to slow down
the speed of WAL urgently, but...

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-06 12:29  Amit Kapila <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Amit Kapila @ 2018-04-06 12:29 UTC (permalink / raw)
  To: Kyotaro HORIGUCHI <[email protected]>; +Cc: [email protected]; Dilip Kumar <[email protected]>; pgsql-hackers

On Fri, Apr 6, 2018 at 1:50 PM, Kyotaro HORIGUCHI
<[email protected]> wrote:
> Hello.
>
> At Wed, 04 Apr 2018 17:26:46 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
>> > In general, I was wondering why in the first place this variable
>> > (full_page_writes) is a SIGHUP variable?  I think if the user tries to
>> > switch it to 'on' from 'off', it won't guarantee the recovery from
>> > torn pages.  Yeah, one can turn it to 'off' from 'on' without any
>> > problem, but as the reverse doesn't guarantee anything, it can confuse
>> > users. What do you think?
>>
>> I tend to agree with you. It works as expected after the next
>> checkpoint. So define the variable as "it can be changed any time
>> but has an effect at the next checkpoint time", then remove
>> XLOG_FPW_CHANGE record. And that eliminates the problem of
>> concurrent calls since the checkpointer becomes the only modifier
>> of the variable. And the problematic function
>> UpdateFullPageWrites also no longer needs to write a WAL
>> record. The information is conveyed only by checkpoint records.
>
> I noticed that XLOG_FPW_CHANGE(fpw=false) is still required by
> pg_start/stop_backup to know FPW's turning-off without waiting
> for the next checkpoint record. But XLOG_FPW_CHANGE(true) is not
> required since no one uses the information. It seems even harmful
> when it is written at the incorrect place.
>
> In the attached patch, shared fullPageWrites is updated only at
> REDO point
>

I am not completely sure if that is the right option because this
would mean that we are defining the new scope for a GUC variable.  I
guess we should take the input of others as well.  I am not sure what
is the right way to do that, but maybe we can start a new thread with
a proper subject and description rather than discussing this under
some related bug fix patch discussion.  I guess we can try that after
CF unless some other people pitch in and share their feedback.


-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-09 08:13  Kyotaro HORIGUCHI <[email protected]>
  parent: Amit Kapila <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-09 08:13 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; pgsql-hackers


At Fri, 6 Apr 2018 17:59:58 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1+1zULC52G_EyNcrrxFCmBi4NUuA1CoQAKu2FFPai_Teg@mail.gmail.com>
> On Fri, Apr 6, 2018 at 1:50 PM, Kyotaro HORIGUCHI
> <[email protected]> wrote:
> > Hello.
> >
> > At Wed, 04 Apr 2018 17:26:46 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
> >> > In general, I was wondering why in the first place this variable
> >> > (full_page_writes) is a SIGHUP variable?  I think if the user tries to
> >> > switch it to 'on' from 'off', it won't guarantee the recovery from
> >> > torn pages.  Yeah, one can turn it to 'off' from 'on' without any
> >> > problem, but as the reverse doesn't guarantee anything, it can confuse
> >> > users. What do you think?
> >>
> >> I tend to agree with you. It works as expected after the next
> >> checkpoint. So define the variable as "it can be changed any time
> >> but has an effect at the next checkpoint time", then remove
> >> XLOG_FPW_CHANGE record. And that eliminates the problem of
> >> concurrent calls since the checkpointer becomes the only modifier
> >> of the variable. And the problematic function
> >> UpdateFullPageWrites also no longer needs to write a WAL
> >> record. The information is conveyed only by checkpoint records.
> >
> > I noticed that XLOG_FPW_CHANGE(fpw=false) is still required by
> > pg_start/stop_backup to know FPW's turning-off without waiting
> > for the next checkpoint record. But XLOG_FPW_CHANGE(true) is not
> > required since no one uses the information. It seems even harmful
> > when it is written at the incorrect place.
> >
> > In the attached patch, shared fullPageWrites is updated only at
> > REDO point
> >
> 
> I am not completely sure if that is the right option because this
> would mean that we are defining the new scope for a GUC variable.  I
> guess we should take the input of others as well.  I am not sure what
> is the right way to do that, but maybe we can start a new thread with
> a proper subject and description rather than discussing this under
> some related bug fix patch discussion.  I guess we can try that after
> CF unless some other people pitch in and share their feedback.

I'd like to refrain from making a new thread since this topic is
registered as an open item (in Old Bugs section). Or making a new
thread then relinking it from the page is preferable?

I'm surprised a bit that this got confilcted so soon. On the way
rebasing, for anyone's information, I considered comment and
documentation fix but all comments and documentation can be read
as both old and new behavior. That being said, the patch contains
a small addtion about the new behavior.

regards,

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-11 11:09  Heikki Linnakangas <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 2 replies; 292+ messages in thread

From: Heikki Linnakangas @ 2018-04-11 11:09 UTC (permalink / raw)
  To: Kyotaro HORIGUCHI <[email protected]>; [email protected]; +Cc: [email protected]; [email protected]; pgsql-hackers

On 09/04/18 11:13, Kyotaro HORIGUCHI wrote:
> 
> At Fri, 6 Apr 2018 17:59:58 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1+1zULC52G_EyNcrrxFCmBi4NUuA1CoQAKu2FFPai_Teg@mail.gmail.com>
>> On Fri, Apr 6, 2018 at 1:50 PM, Kyotaro HORIGUCHI
>> <[email protected]> wrote:
>>> Hello.
>>>
>>> At Wed, 04 Apr 2018 17:26:46 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
>>>>> In general, I was wondering why in the first place this variable
>>>>> (full_page_writes) is a SIGHUP variable?  I think if the user tries to
>>>>> switch it to 'on' from 'off', it won't guarantee the recovery from
>>>>> torn pages.  Yeah, one can turn it to 'off' from 'on' without any
>>>>> problem, but as the reverse doesn't guarantee anything, it can confuse
>>>>> users. What do you think?
>>>>
>>>> I tend to agree with you. It works as expected after the next
>>>> checkpoint. So define the variable as "it can be changed any time
>>>> but has an effect at the next checkpoint time", then remove
>>>> XLOG_FPW_CHANGE record. And that eliminates the problem of
>>>> concurrent calls since the checkpointer becomes the only modifier
>>>> of the variable. And the problematic function
>>>> UpdateFullPageWrites also no longer needs to write a WAL
>>>> record. The information is conveyed only by checkpoint records.
>>>
>>> I noticed that XLOG_FPW_CHANGE(fpw=false) is still required by
>>> pg_start/stop_backup to know FPW's turning-off without waiting
>>> for the next checkpoint record. But XLOG_FPW_CHANGE(true) is not
>>> required since no one uses the information. It seems even harmful
>>> when it is written at the incorrect place.
>>>
>>> In the attached patch, shared fullPageWrites is updated only at
>>> REDO point
>>
>> I am not completely sure if that is the right option because this
>> would mean that we are defining the new scope for a GUC variable.  I
>> guess we should take the input of others as well.  I am not sure what
>> is the right way to do that, but maybe we can start a new thread with
>> a proper subject and description rather than discussing this under
>> some related bug fix patch discussion.  I guess we can try that after
>> CF unless some other people pitch in and share their feedback.

I think the new behavior where the GUC only takes effect at next 
checkpoint is OK. It seems quite intuitive.

> [rebased patch version]

Looks good at a quick glance. Assuming no objections from others, I'll 
take a closer look and commit tomorrow. Thanks!

- Heikki




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-11 20:24  Michael Paquier <[email protected]>
  parent: Heikki Linnakangas <[email protected]>
  1 sibling, 1 reply; 292+ messages in thread

From: Michael Paquier @ 2018-04-11 20:24 UTC (permalink / raw)
  To: Heikki Linnakangas <[email protected]>; +Cc: Kyotaro HORIGUCHI <[email protected]>; [email protected]; [email protected]; pgsql-hackers

On Wed, Apr 11, 2018 at 02:09:48PM +0300, Heikki Linnakangas wrote:
> I think the new behavior where the GUC only takes effect at next checkpoint
> is OK. It seems quite intuitive.
> 
> > [rebased patch version]
> 
> Looks good at a quick glance. Assuming no objections from others, I'll take
> a closer look and commit tomorrow. Thanks!

Sorry for not following up closely this thread lately.

+   /*
+    * If full_page_writes has been turned off, issue XLOG_FPW_CHANGE before
+    * the flag actually takes effect. No lock is required since checkpointer
+    * is the only updator of shared fullPageWrites after recovery is
+    * finished. Both shared and local fullPageWrites do not change before the
+    * next reading below.
+    */
+   if (Insert->fullPageWrites && !fullPageWrites)
+   {
+       XLogBeginInsert();
+       XLogRegisterData((char *) (&fullPageWrites), sizeof(bool));
+       XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
+   }

This is not actually true.  If a fallback_promote is used, then
CreateCheckPoint() is called by the startup process which is in charge
of issuing the end-of-recovery checkpoint, and not the checkpointer.  So
I still fail to see how a no-lock approach is fine except if we remove
fallback_promote?
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-12 01:34  Kyotaro HORIGUCHI <[email protected]>
  parent: Michael Paquier <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-12 01:34 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; pgsql-hackers

Hello. Thanks to Heikkit for picking this up and thanks for the
commnet to Michael.

# The attached is changed only in a comment, and rebased.

At Thu, 12 Apr 2018 05:24:14 +0900, Michael Paquier <[email protected]> wrote in <[email protected]>
> On Wed, Apr 11, 2018 at 02:09:48PM +0300, Heikki Linnakangas wrote:
> > I think the new behavior where the GUC only takes effect at next checkpoint
> > is OK. It seems quite intuitive.
> > 
> > > [rebased patch version]
> > 
> > Looks good at a quick glance. Assuming no objections from others, I'll take
> > a closer look and commit tomorrow. Thanks!
> 
> Sorry for not following up closely this thread lately.
> 
> +   /*
> +    * If full_page_writes has been turned off, issue XLOG_FPW_CHANGE before
> +    * the flag actually takes effect. No lock is required since checkpointer
> +    * is the only updator of shared fullPageWrites after recovery is
> +    * finished. Both shared and local fullPageWrites do not change before the
> +    * next reading below.
> +    */
> +   if (Insert->fullPageWrites && !fullPageWrites)
> +   {
> +       XLogBeginInsert();
> +       XLogRegisterData((char *) (&fullPageWrites), sizeof(bool));
> +       XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
> +   }
> 
> This is not actually true.  If a fallback_promote is used, then
> CreateCheckPoint() is called by the startup process which is in charge
> of issuing the end-of-recovery checkpoint, and not the checkpointer.  So
> I still fail to see how a no-lock approach is fine except if we remove
> fallback_promote?

Checkpointer never calls CreateCheckPoint while
RecoveryInProgress() == true. In other words, checkpointer is not
an updator of shared FPW at the time StartupXLOG calls
CreateCheckPoint for fallback_promote.

The comment may be somewhat confusing that it is written
there. The point is that checkpointer and StartupXLOG are
mutually excluded on updating shared FPW by
SharedRecoveryInProgress flag.

| * If full_page_writes has been turned off, issue XLOG_FPW_CHANGE before
| * the flag actually takes effect. Checkpointer never calls this function
| * before StartupXLOG() turns off SharedRecoveryInProgress so there's no
| * window where checkpointer and startup processes - the only updators of
| * the flag - can update shared FPW simultaneously. Thus no lock is
| * required here. Both shared and local fullPageWrites do not change
| * before the next reading below.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-12 05:07  Michael Paquier <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Michael Paquier @ 2018-04-12 05:07 UTC (permalink / raw)
  To: Kyotaro HORIGUCHI <[email protected]>; +Cc: [email protected]; [email protected]; [email protected]; pgsql-hackers

On Thu, Apr 12, 2018 at 10:34:30AM +0900, Kyotaro HORIGUCHI wrote:
> Checkpointer never calls CreateCheckPoint while
> RecoveryInProgress() == true. In other words, checkpointer is not
> an updator of shared FPW at the time StartupXLOG calls
> CreateCheckPoint for fallback_promote.

I have been able to spend a couple of hours on your patch, wrapping my
mind on your stuff.  So what I had in mind was something like this type
of scenario:
1) The startup process requires a restart point.
2) The checkpointer receives the request, and blocks before reading
RecoveryInProgress().
3) A fallback_promote is triggered, making the startup process call
CreateCheckpoint().
4) Startup process finishes checkpoint, updates Insert->fullPageWrites.
5) Checkpoint reads RecoveryInProgress to false, moves on with its
checkpoint.

> The comment may be somewhat confusing that it is written
> there. The point is that checkpointer and StartupXLOG are
> mutually excluded on updating shared FPW by
> SharedRecoveryInProgress flag.

Indeed.  I can see that it is the main key point of the patch.

> | * If full_page_writes has been turned off, issue XLOG_FPW_CHANGE before
> | * the flag actually takes effect. Checkpointer never calls this function
> | * before StartupXLOG() turns off SharedRecoveryInProgress so there's no
> | * window where checkpointer and startup processes - the only updators of
> | * the flag - can update shared FPW simultaneously. Thus no lock is
> | * required here. Both shared and local fullPageWrites do not change
> | * before the next reading below.

Yeah, this reduces the confusion.

(The latest patch is a mix of two patches.)

+        The default is <literal>on</literal>. The change of the parmeter takes
+        effect at the next checkpoint time.
s/parmeter/parameter/

By the way, I would vote for keeping track in WAL of full_page_writes
switched from off to on.  This is not used in the backend, but that's
still useful for debugging end-user issues.

Actually, I was wondering why a spin lock is not taken in
RecoveryInProgress when reading SharedRecoveryInProgress, but that's
from 1a3d1044 which added a memory barrier instead as guarantee...
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-12 07:59  Kyotaro HORIGUCHI <[email protected]>
  parent: Michael Paquier <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-12 07:59 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; pgsql-hackers

Hello.

At Thu, 12 Apr 2018 14:07:53 +0900, Michael Paquier <[email protected]> wrote in <[email protected]>
> On Thu, Apr 12, 2018 at 10:34:30AM +0900, Kyotaro HORIGUCHI wrote:
> > Checkpointer never calls CreateCheckPoint while
> > RecoveryInProgress() == true. In other words, checkpointer is not
> > an updator of shared FPW at the time StartupXLOG calls
> > CreateCheckPoint for fallback_promote.
> 
> I have been able to spend a couple of hours on your patch, wrapping my
> mind on your stuff.  So what I had in mind was something like this type
> of scenario:

Thank for the precise explanation.

The scenario that CreateCheckPoint is called simultaneously in
different prcoesses seems broken by itself in the first place but
I put that aside for now.

> 1) The startup process requires a restart point.
> 2) The checkpointer receives the request, and blocks before reading
> RecoveryInProgress().

RecoveryInProgress doesn't take lock. But I assume here that
checkpointer is taking a long time after entering
RecoveryInProgress and haven't actually read
SharedRecoveryInProgress.

> 3) A fallback_promote is triggered, making the startup process call
> CreateCheckpoint().

I'm confused here. It seems to me that StartupXLOG calls
CreateCheckPoint only in bootstrap or standalone cases. No
concurrent processe is running in the cases.

Even if CreateCheckPoint is called there in the IsUnderPostmater
case, checkpointer never calls CreateCheckPoint during the
checkpoint. This is safe in regard to shared FPW. (but checkpoint
is being blocked in this scenario.)

Assuming that RequestCheckpoint() is called here, the request is
merged with the previous request above and the function returns
after the checkpoint ends. (That is, checkpointer must continue
to run in this case.)

> 4) Startup process finishes checkpoint, updates Insert->fullPageWrites.

According to this scenario, checkpionter is still stalling
now. So it is not a concurrent update.

> 5) Checkpoint reads RecoveryInProgress to false, moves on with its
> checkpoint.

If checkpointer sees SharedRecoveryInProgress being false,
Create(or Request)CheckPoint called in (3) must have finished and
StartupXLOG() no longer updates shared FPW. There's no concurrent
update.

> > The comment may be somewhat confusing that it is written
> > there. The point is that checkpointer and StartupXLOG are
> > mutually excluded on updating shared FPW by
> > SharedRecoveryInProgress flag.
> 
> Indeed.  I can see that it is the main key point of the patch.
> 
> > | * If full_page_writes has been turned off, issue XLOG_FPW_CHANGE before
> > | * the flag actually takes effect. Checkpointer never calls this function
> > | * before StartupXLOG() turns off SharedRecoveryInProgress so there's no
> > | * window where checkpointer and startup processes - the only updators of
> > | * the flag - can update shared FPW simultaneously. Thus no lock is
> > | * required here. Both shared and local fullPageWrites do not change
> > | * before the next reading below.
> 
> Yeah, this reduces the confusion.

Thanks^^;

> (The latest patch is a mix of two patches.)

Sorry I counld get this.

> +        The default is <literal>on</literal>. The change of the parmeter takes
> +        effect at the next checkpoint time.
> s/parmeter/parameter/
> 
> By the way, I would vote for keeping track in WAL of full_page_writes
> switched from off to on.  This is not used in the backend, but that's
> still useful for debugging end-user issues.

Agreed and I tried that. The problem on that is that some records
can be written after REDO point before XLOG_FPW_CHANGE(true) is
written. However this is no problem for the FPW-related stuff to
work properly (since no one looks it), the FPW record suggests
that the current checkpoint loses FPI in the first several
records. This has a far larger impact with this patch because
shared FPW is always turned on just at REDO point.

So I choosed not to write XLOG_FPW_CHANGE(false) rather than
writing bogus records.

> Actually, I was wondering why a spin lock is not taken in
> RecoveryInProgress when reading SharedRecoveryInProgress, but that's
> from 1a3d1044 which added a memory barrier instead as guarantee...

Maybe it doesn't need barrier, since the flag is initialized as
true and becomes false just once and delay in reading by other
processes doesn't no harm. I think that bool doesn't suffer
atomicity. Even all these are true, some description is needed.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-12 08:10  Michael Paquier <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Michael Paquier @ 2018-04-12 08:10 UTC (permalink / raw)
  To: Kyotaro HORIGUCHI <[email protected]>; +Cc: [email protected]; [email protected]; [email protected]; pgsql-hackers

On Thu, Apr 12, 2018 at 04:59:10PM +0900, Kyotaro HORIGUCHI wrote:
> At Thu, 12 Apr 2018 14:07:53 +0900, Michael Paquier <[email protected]> wrote in <[email protected]>
>> I have been able to spend a couple of hours on your patch, wrapping my
>> mind on your stuff.  So what I had in mind was something like this type
>> of scenario:
> 
> Thank for the precise explanation.

Just to be clear and to avoid incorrect conclusion.  This is the type of
scenarios I imagined about when I read your previous email, concluding
such scenarios those cannot apply per the strong assumption on
SharedRecoveryInProgress your patch heavily relies on.  In short I have
no objections.

>> (The latest patch is a mix of two patches.)
> 
> Sorry I counld get this.

The patch called v2-0001-Change-FPW-handling.patch posted on
https://www.postgresql.org/message-id/20180412.103430.133595350.horiguchi.kyotaro%40lab.ntt.co.jp,
which is the latest version available, is a mix of the patch you are
creating for this thread and of a patch aimed a fixing an issue with
partition range handling.
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-12 18:55  Robert Haas <[email protected]>
  parent: Heikki Linnakangas <[email protected]>
  1 sibling, 1 reply; 292+ messages in thread

From: Robert Haas @ 2018-04-12 18:55 UTC (permalink / raw)
  To: Heikki Linnakangas <[email protected]>; +Cc: Kyotaro HORIGUCHI <[email protected]>; Amit Kapila <[email protected]>; Michael Paquier <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Wed, Apr 11, 2018 at 7:09 AM, Heikki Linnakangas <[email protected]> wrote:
> I think the new behavior where the GUC only takes effect at next checkpoint
> is OK. It seems quite intuitive.

I think it may actually be confusing.  If you run pg_ctl reload and it
reports that the value has changed, you'll expect it to have taken
effect.  But really, it will take effect at some later time.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-13 01:29  Michael Paquier <[email protected]>
  parent: Robert Haas <[email protected]>
  0 siblings, 2 replies; 292+ messages in thread

From: Michael Paquier @ 2018-04-13 01:29 UTC (permalink / raw)
  To: Robert Haas <[email protected]>; +Cc: Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Amit Kapila <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Thu, Apr 12, 2018 at 02:55:53PM -0400, Robert Haas wrote:
> I think it may actually be confusing.  If you run pg_ctl reload and it
> reports that the value has changed, you'll expect it to have taken
> effect.  But really, it will take effect at some later time.

It is true that sometimes some people like to temporarily disable
full_page_writes particularly when doing some bulk load of data to
minimize the effort on WAL, and then re-enable it just after doing 
the inserting this data.

Still does it matter when the change is effective?  By disabling
full_page_writes even temporarily, you accept the fact that this
instance would not be safe until the next checkpoint completes.  The
instance even finishes by writing less unnecessary WAL data if the
change is only effective at the next checkpoint.  Well, it is true that
this increases potential torn pages problems but the user is already
accepting that risk if a crash happens until the next checkpoint then it
exposes itself to torn pages anyway as it chose to disable
full_page_writes.
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-13 03:01  Amit Kapila <[email protected]>
  parent: Michael Paquier <[email protected]>
  1 sibling, 1 reply; 292+ messages in thread

From: Amit Kapila @ 2018-04-13 03:01 UTC (permalink / raw)
  To: Michael Paquier <[email protected]>; +Cc: Robert Haas <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Fri, Apr 13, 2018 at 6:59 AM, Michael Paquier <[email protected]> wrote:
> On Thu, Apr 12, 2018 at 02:55:53PM -0400, Robert Haas wrote:
>> I think it may actually be confusing.  If you run pg_ctl reload and it
>> reports that the value has changed, you'll expect it to have taken
>> effect.  But really, it will take effect at some later time.
>

+1. I also think it is confusing and it could be difficult for end
users to know when the setting is effective.

> It is true that sometimes some people like to temporarily disable
> full_page_writes particularly when doing some bulk load of data to
> minimize the effort on WAL, and then re-enable it just after doing
> the inserting this data.
>
> Still does it matter when the change is effective?  By disabling
> full_page_writes even temporarily, you accept the fact that this
> instance would not be safe until the next checkpoint completes.  The
> instance even finishes by writing less unnecessary WAL data if the
> change is only effective at the next checkpoint.  Well, it is true that
> this increases potential torn pages problems but the user is already
> accepting that risk if a crash happens until the next checkpoint then it
> exposes itself to torn pages anyway as it chose to disable
> full_page_writes.
>

I think this means that is will be difficult for end users to predict
unless they track the next checkpoint which isn't too bad, but won't
be convenient either.

-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-13 04:47  Kyotaro HORIGUCHI <[email protected]>
  parent: Amit Kapila <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-13 04:47 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; [email protected]; pgsql-hackers

At Fri, 13 Apr 2018 08:31:02 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1LVFpLf=d-7XmfwhLv7Xu53pU0bGU=wVrYWSRU4XSsyHQ@mail.gmail.com>
> On Fri, Apr 13, 2018 at 6:59 AM, Michael Paquier <[email protected]> wrote:
> > On Thu, Apr 12, 2018 at 02:55:53PM -0400, Robert Haas wrote:
> >> I think it may actually be confusing.  If you run pg_ctl reload and it
> >> reports that the value has changed, you'll expect it to have taken
> >> effect.  But really, it will take effect at some later time.
> >
> 
> +1. I also think it is confusing and it could be difficult for end
> users to know when the setting is effective.
> 
> > It is true that sometimes some people like to temporarily disable
> > full_page_writes particularly when doing some bulk load of data to
> > minimize the effort on WAL, and then re-enable it just after doing
> > the inserting this data.
> >
> > Still does it matter when the change is effective?  By disabling
> > full_page_writes even temporarily, you accept the fact that this
> > instance would not be safe until the next checkpoint completes.  The
> > instance even finishes by writing less unnecessary WAL data if the
> > change is only effective at the next checkpoint.  Well, it is true that
> > this increases potential torn pages problems but the user is already
> > accepting that risk if a crash happens until the next checkpoint then it
> > exposes itself to torn pages anyway as it chose to disable
> > full_page_writes.

I still don't think that enabling FPW anytime is useful but
disabling seems useful as I mentioned upthread.

The problem was checkpointer changes the flag anytime including
recovery time. Startup process updates the same flag at the end
of recovery but before publicated. Letting checkpointer change
the flag only at checkpoint time is a straightforward way to
avoid conflicts with startup process.

I reconsider a bit and came up with the thought that we could
just skip changing shared FPW in checkpointer until recovery
ends, then update the flag after recovery end (perhaps at
checkpoint time in major cases). In this case, FPI is attached
from REDO point of the first checkpoint (not restartpoint) or a
bit earlier, then FPW can be flipped at any time.  I'll come up
with that soon.

> I think this means that is will be difficult for end users to predict
> unless they track the next checkpoint which isn't too bad, but won't
> be convenient either.

Looking checkpiont record is enough to know wheter the checkpoint
is protected by FPW eough, but I agree that such strictness is
not crutial.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-13 08:28  Kyotaro HORIGUCHI <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-13 08:28 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; [email protected]; pgsql-hackers

At Fri, 13 Apr 2018 13:47:51 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
> At Fri, 13 Apr 2018 08:31:02 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1LVFpLf=d-7XmfwhLv7Xu53pU0bGU=wVrYWSRU4XSsyHQ@mail.gmail.com>
> > On Fri, Apr 13, 2018 at 6:59 AM, Michael Paquier <[email protected]> wrote:
> > > On Thu, Apr 12, 2018 at 02:55:53PM -0400, Robert Haas wrote:
> > >> I think it may actually be confusing.  If you run pg_ctl reload and it
> > >> reports that the value has changed, you'll expect it to have taken
> > >> effect.  But really, it will take effect at some later time.
> > >
> > 
> > +1. I also think it is confusing and it could be difficult for end
> > users to know when the setting is effective.
> > 
> > > It is true that sometimes some people like to temporarily disable
> > > full_page_writes particularly when doing some bulk load of data to
> > > minimize the effort on WAL, and then re-enable it just after doing
> > > the inserting this data.
> > >
> > > Still does it matter when the change is effective?  By disabling
> > > full_page_writes even temporarily, you accept the fact that this
> > > instance would not be safe until the next checkpoint completes.  The
> > > instance even finishes by writing less unnecessary WAL data if the
> > > change is only effective at the next checkpoint.  Well, it is true that
> > > this increases potential torn pages problems but the user is already
> > > accepting that risk if a crash happens until the next checkpoint then it
> > > exposes itself to torn pages anyway as it chose to disable
> > > full_page_writes.
> 
> I still don't think that enabling FPW anytime is useful but
> disabling seems useful as I mentioned upthread.
> 
> The problem was checkpointer changes the flag anytime including
> recovery time. Startup process updates the same flag at the end
> of recovery but before publicated. Letting checkpointer change
> the flag only at checkpoint time is a straightforward way to
> avoid conflicts with startup process.
> 
> I reconsider a bit and came up with the thought that we could
> just skip changing shared FPW in checkpointer until recovery
> ends, then update the flag after recovery end (perhaps at
> checkpoint time in major cases). In this case, FPI is attached
> from REDO point of the first checkpoint (not restartpoint) or a
> bit earlier, then FPW can be flipped at any time.  I'll come up
> with that soon.

Please find the attached. The most significant change is that
UpdateSharedMemoryConfig skips updating of shared fullPageWrites
during recovery. The original crash is fixed since this
guarantees that XLog working area is initializeed before reaching
UpdateFullPageWrites().

Addition to that, I changed CheckpointerMain so that it tries
update of shared FPW regardless of SIGHUP and provided new
function to just wakeup checkpointer. StartupXLOG wakes up
checkpointer either checkpoint is required or not and
checkpointer makes the first update of shared FPW at the time.

After this point, everything works as the same as the current
behavior.

> > I think this means that is will be difficult for end users to predict
> > unless they track the next checkpoint which isn't too bad, but won't
> > be convenient either.
> 
> Looking checkpiont record is enough to know wheter the checkpoint
> is protected by FPW eough, but I agree that such strictness is
> not crutial.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-13 08:34  Kyotaro HORIGUCHI <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-13 08:34 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; [email protected]; pgsql-hackers

Sorry, the patch attached to the previous main is slightly
old. The attached is the correct one.

# They differ only in some phrase in a comment.

====
At Fri, 13 Apr 2018 17:28:40 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
At Fri, 13 Apr 2018 13:47:51 +0900 (Tokyo Standard Time), Kyotaro HORIGUCHI <[email protected]> wrote in <[email protected]>
> At Fri, 13 Apr 2018 08:31:02 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1LVFpLf=d-7XmfwhLv7Xu53pU0bGU=wVrYWSRU4XSsyHQ@mail.gmail.com>
> > On Fri, Apr 13, 2018 at 6:59 AM, Michael Paquier <[email protected]> wrote:
> > > On Thu, Apr 12, 2018 at 02:55:53PM -0400, Robert Haas wrote:
> > >> I think it may actually be confusing.  If you run pg_ctl reload and it
> > >> reports that the value has changed, you'll expect it to have taken
> > >> effect.  But really, it will take effect at some later time.
> > >
> > 
> > +1. I also think it is confusing and it could be difficult for end
> > users to know when the setting is effective.
> > 
> > > It is true that sometimes some people like to temporarily disable
> > > full_page_writes particularly when doing some bulk load of data to
> > > minimize the effort on WAL, and then re-enable it just after doing
> > > the inserting this data.
> > >
> > > Still does it matter when the change is effective?  By disabling
> > > full_page_writes even temporarily, you accept the fact that this
> > > instance would not be safe until the next checkpoint completes.  The
> > > instance even finishes by writing less unnecessary WAL data if the
> > > change is only effective at the next checkpoint.  Well, it is true that
> > > this increases potential torn pages problems but the user is already
> > > accepting that risk if a crash happens until the next checkpoint then it
> > > exposes itself to torn pages anyway as it chose to disable
> > > full_page_writes.
> 
> I still don't think that enabling FPW anytime is useful but
> disabling seems useful as I mentioned upthread.
> 
> The problem was checkpointer changes the flag anytime including
> recovery time. Startup process updates the same flag at the end
> of recovery but before publicated. Letting checkpointer change
> the flag only at checkpoint time is a straightforward way to
> avoid conflicts with startup process.
> 
> I reconsider a bit and came up with the thought that we could
> just skip changing shared FPW in checkpointer until recovery
> ends, then update the flag after recovery end (perhaps at
> checkpoint time in major cases). In this case, FPI is attached
> from REDO point of the first checkpoint (not restartpoint) or a
> bit earlier, then FPW can be flipped at any time.  I'll come up
> with that soon.

Please find the attached. The most significant change is that
UpdateSharedMemoryConfig skips updating of shared fullPageWrites
during recovery. The original crash is fixed since this
guarantees that XLog working area is initializeed before reaching
UpdateFullPageWrites().

Addition to that, I changed CheckpointerMain so that it tries
update of shared FPW regardless of SIGHUP and provided new
function to just wakeup checkpointer. StartupXLOG wakes up
checkpointer either checkpoint is required or not and
checkpointer makes the first update of shared FPW at the time.

After this point, everything works as the same as the current
behavior.

> > I think this means that is will be difficult for end users to predict
> > unless they track the next checkpoint which isn't too bad, but won't
> > be convenient either.
> 
> Looking checkpiont record is enough to know wheter the checkpoint
> is protected by FPW eough, but I agree that such strictness is
> not crutial.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-13 17:06  Robert Haas <[email protected]>
  parent: Michael Paquier <[email protected]>
  1 sibling, 1 reply; 292+ messages in thread

From: Robert Haas @ 2018-04-13 17:06 UTC (permalink / raw)
  To: Michael Paquier <[email protected]>; +Cc: Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Amit Kapila <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Thu, Apr 12, 2018 at 9:29 PM, Michael Paquier <[email protected]> wrote:
> Still does it matter when the change is effective?

I don't really care deeply about when the change takes effect, but I
do care about whether the time when the system *says* the change took
effect is the same as when it *actually* took effect.  If those aren't
the same, it's confusing.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-18 14:37  Amit Kapila <[email protected]>
  parent: Robert Haas <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Amit Kapila @ 2018-04-18 14:37 UTC (permalink / raw)
  To: Robert Haas <[email protected]>; +Cc: Michael Paquier <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Fri, Apr 13, 2018 at 10:36 PM, Robert Haas <[email protected]> wrote:
> On Thu, Apr 12, 2018 at 9:29 PM, Michael Paquier <[email protected]> wrote:
>> Still does it matter when the change is effective?
>
> I don't really care deeply about when the change takes effect, but I
> do care about whether the time when the system *says* the change took
> effect is the same as when it *actually* took effect.  If those aren't
> the same, it's confusing.
>

So, what in your opinion is the way to deal with this?  If we make it
a PGC_POSTMASTER parameter, it will have a very clear behavior and
users don't need to bother whether they have a risk of torn page
problem or not and as a side-impact the code will be simplified as
well.  However, as Michael said the people who get the benefit of this
option by disabling/enabling this parameter might complain.  Keeping
it as a SIGHUP option has the drawback that even after the user has
enabled it, there is a risk of torn pages.

-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-18 14:52  Robert Haas <[email protected]>
  parent: Amit Kapila <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Robert Haas @ 2018-04-18 14:52 UTC (permalink / raw)
  To: Amit Kapila <[email protected]>; +Cc: Michael Paquier <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Wed, Apr 18, 2018 at 10:37 AM, Amit Kapila <[email protected]> wrote:
> On Fri, Apr 13, 2018 at 10:36 PM, Robert Haas <[email protected]> wrote:
>> On Thu, Apr 12, 2018 at 9:29 PM, Michael Paquier <[email protected]> wrote:
>>> Still does it matter when the change is effective?
>>
>> I don't really care deeply about when the change takes effect, but I
>> do care about whether the time when the system *says* the change took
>> effect is the same as when it *actually* took effect.  If those aren't
>> the same, it's confusing.
>>
>
> So, what in your opinion is the way to deal with this?  If we make it
> a PGC_POSTMASTER parameter, it will have a very clear behavior and
> users don't need to bother whether they have a risk of torn page
> problem or not and as a side-impact the code will be simplified as
> well.  However, as Michael said the people who get the benefit of this
> option by disabling/enabling this parameter might complain.  Keeping
> it as a SIGHUP option has the drawback that even after the user has
> enabled it, there is a risk of torn pages.

I would just document the risks.  If the documentation says that you
can't rely on the value until after the next checkpoint, or whatever
the rule is, then I think we're fine.  I don't think that we really
have the infrastructure to do any better; if we try, we'll just end up
with odd warts.  Documenting the current set of warts is less churn
and less work.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-19 01:49  Michael Paquier <[email protected]>
  parent: Robert Haas <[email protected]>
  0 siblings, 3 replies; 292+ messages in thread

From: Michael Paquier @ 2018-04-19 01:49 UTC (permalink / raw)
  To: Robert Haas <[email protected]>; +Cc: Amit Kapila <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Wed, Apr 18, 2018 at 10:52:51AM -0400, Robert Haas wrote:
> I would just document the risks.  If the documentation says that you
> can't rely on the value until after the next checkpoint, or whatever
> the rule is, then I think we're fine.  I don't think that we really
> have the infrastructure to do any better; if we try, we'll just end up
> with odd warts.  Documenting the current set of warts is less churn
> and less work.

The last version of the patch proposed has eaten this diff which was
part of one of the past versions (v2-0001-Change-FPW-handling.patch from
https://www.postgresql.org/message-id/20180412.103430.133595350.horiguchi.kyotaro%40lab.ntt.co.jp):
+        The default is <literal>on</literal>. The change of the parameter takes
+        effect at the next checkpoint time.
So there were some documentation about the beHavior change for what it's
worth. 

And, er, actually, I was thinking again about the case where a user
wants to disable full_page_writes temporarily to do some bulk load and
then re-enable it.  With the patch proposed to actually update the FPW
effect at checkpoint time, then a user would need to issue a manual
checkpoint after updating the configuration and reloading, which may
create more I/O than he'd want to pay for, then a second checkpoint
would need to be issued after the configuration comes back again.  That
would cause a regression which could surprise a class of users.  WAL and
FPW overhead is a problem which shows up a lot when doing bulk-loading
of data.
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-19 13:41  Amit Kapila <[email protected]>
  parent: Michael Paquier <[email protected]>
  2 siblings, 2 replies; 292+ messages in thread

From: Amit Kapila @ 2018-04-19 13:41 UTC (permalink / raw)
  To: Michael Paquier <[email protected]>; +Cc: Robert Haas <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Thu, Apr 19, 2018 at 7:19 AM, Michael Paquier <[email protected]> wrote:
> On Wed, Apr 18, 2018 at 10:52:51AM -0400, Robert Haas wrote:
>> I would just document the risks.  If the documentation says that you
>> can't rely on the value until after the next checkpoint, or whatever
>> the rule is, then I think we're fine.  I don't think that we really
>> have the infrastructure to do any better; if we try, we'll just end up
>> with odd warts.  Documenting the current set of warts is less churn
>> and less work.
>
> The last version of the patch proposed has eaten this diff which was
> part of one of the past versions (v2-0001-Change-FPW-handling.patch from
> https://www.postgresql.org/message-id/20180412.103430.133595350.horiguchi.kyotaro%40lab.ntt.co.jp):
> +        The default is <literal>on</literal>. The change of the parameter takes
> +        effect at the next checkpoint time.
> So there were some documentation about the beHavior change for what it's
> worth.
>
> And, er, actually, I was thinking again about the case where a user
> wants to disable full_page_writes temporarily to do some bulk load and
> then re-enable it.  With the patch proposed to actually update the FPW
> effect at checkpoint time, then a user would need to issue a manual
> checkpoint after updating the configuration and reloading, which may
> create more I/O than he'd want to pay for, then a second checkpoint
> would need to be issued after the configuration comes back again.
>

Why a second checkpoint?  One checkpoint either manual or automatic
should be enough to make the setting effective.

-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-20 01:04  Michael Paquier <[email protected]>
  parent: Amit Kapila <[email protected]>
  1 sibling, 0 replies; 292+ messages in thread

From: Michael Paquier @ 2018-04-20 01:04 UTC (permalink / raw)
  To: Amit Kapila <[email protected]>; +Cc: Robert Haas <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Thu, Apr 19, 2018 at 07:11:43PM +0530, Amit Kapila wrote:
> On Thu, Apr 19, 2018 at 7:19 AM, Michael Paquier <[email protected]> wrote:
>> And, er, actually, I was thinking again about the case where a user
>> wants to disable full_page_writes temporarily to do some bulk load and
>> then re-enable it.  With the patch proposed to actually update the FPW
>> effect at checkpoint time, then a user would need to issue a manual
>> checkpoint after updating the configuration and reloading, which may
>> create more I/O than he'd want to pay for, then a second checkpoint
>> would need to be issued after the configuration comes back again.
> 
> Why a second checkpoint?  One checkpoint either manual or automatic
> should be enough to make the setting effective.

I was thinking about cases where users have say hourly cron jobs in
charge of doing some maintenance of update cleanups, where they would
need to be sure that full_page_writes are back online after doing the
bulk-load.  In this case an extra checkpoint would be necessary to make
the parameter update effective.
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-20 06:10  Kyotaro HORIGUCHI <[email protected]>
  parent: Michael Paquier <[email protected]>
  2 siblings, 1 reply; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-20 06:10 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; [email protected]; pgsql-hackers

By the way, I think I found a bug of FPW.

The following steps yields INSERT record that doesn't have a FPI
after a checkpoint.

(Start server with full_page_writes = off)
CREATE TABLE t (a int);
CHECKPOINT;
INSERT INTO t VALUES (1);
ALTER SYSTEM SET full_page_writes TO on;
SELECT pg_reload_conf();
CHECKPOINT;
INSERT INTO t VALUES (1);

The last insert is expected to write a record with FPI but it
doesn't actually. No FPI will be written for the page after that.

It seems that the reason is that XLogInsertRecord is forgetting
to check doPageWrites' update.

In the failure case, fpw_lsn has been set by XLogRecordAssemble
but doPageWrite is false at the time and it considers that no FPI
is required then it sets fpw_lsn to InvalidXLogRecPtr.

After that, XLogInsertRecord receives the record but it thinks
that doPageWrites is true since it looks the shared value.

> if (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr && doPageWrites)

So this line thinks that "no FPI is omitted in this record" but
actually the record is just forgotten to attach them.

The attached patch lets XLogInsertRecord check if doPageWrites
has been turned on after the last call and cause recomputation in
the case.

> * If there are any registered buffers, and a full-page image was not taken
> * of all of them, *fpw_lsn is set to the lowest LSN among such pages. This
> * signals that the assembled record is only good for insertion on the
> * assumption that the RedoRecPtr and doPageWrites values were up-to-date.

And the patch fixes one comment typo of XLogInsertRecord.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-20 08:33  Kyotaro HORIGUCHI <[email protected]>
  parent: Amit Kapila <[email protected]>
  1 sibling, 0 replies; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-20 08:33 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; [email protected]; pgsql-hackers

I noticed that the previous patch is a mixture with another
patch.. sorry.

At Thu, 19 Apr 2018 19:11:43 +0530, Amit Kapila <[email protected]> wrote in <CAA4eK1LZ8UBfOMWMHAQGZ0_5N7aWPOXgUYTvCV+J2SXkpmrjRg@mail.gmail.com>
> On Thu, Apr 19, 2018 at 7:19 AM, Michael Paquier <[email protected]> wrote:
> > On Wed, Apr 18, 2018 at 10:52:51AM -0400, Robert Haas wrote:
> >> I would just document the risks.  If the documentation says that you
> >> can't rely on the value until after the next checkpoint, or whatever
> >> the rule is, then I think we're fine.  I don't think that we really
> >> have the infrastructure to do any better; if we try, we'll just end up
> >> with odd warts.  Documenting the current set of warts is less churn
> >> and less work.
> >
> > The last version of the patch proposed has eaten this diff which was
> > part of one of the past versions (v2-0001-Change-FPW-handling.patch from
> > https://www.postgresql.org/message-id/20180412.103430.133595350.horiguchi.kyotaro%40lab.ntt.co.jp):
> > +        The default is <literal>on</literal>. The change of the parameter takes
> > +        effect at the next checkpoint time.
> > So there were some documentation about the beHavior change for what it's
> > worth.
> >
> > And, er, actually, I was thinking again about the case where a user
> > wants to disable full_page_writes temporarily to do some bulk load and
> > then re-enable it.  With the patch proposed to actually update the FPW
> > effect at checkpoint time, then a user would need to issue a manual
> > checkpoint after updating the configuration and reloading, which may
> > create more I/O than he'd want to pay for, then a second checkpoint
> > would need to be issued after the configuration comes back again.
> >
> 
> Why a second checkpoint?  One checkpoint either manual or automatic
> should be enough to make the setting effective.

One significant point in my first patch is anyway the FPW is
useless untill the second checkpoint starts. In the sense of the
timing when *useful* FPW comes back, the second checkpoint is
required regardless of the patch. As Amit said, there is no
difference whether it is manual or automatic.

On the other hand the timing when FPW is actually turned off is
different. Focusing on user's view, one can run bulkload
immediately under the current behavior but should wait for the
next checkpoint starts with the first patch, which can be caused
manually but may be delayed after the currently running
checkpoint if any.

I think that starting the *first* checkpoint before bulkload is
not such a bother but on the other hand the behavior can lead to
FPW flood for those who are accostomed to the current behavior.

The attached first patch is the bugfix proposed in this thread.
The second fixes the cocurrent update problem only.

The third changes the behavior so that turning-on happenes only
on checkpoints and turning-off happens any time.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center


^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-20 12:36  Amit Kapila <[email protected]>
  parent: Kyotaro HORIGUCHI <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Amit Kapila @ 2018-04-20 12:36 UTC (permalink / raw)
  To: Kyotaro HORIGUCHI <[email protected]>; +Cc: Michael Paquier <[email protected]>; Robert Haas <[email protected]>; hlinnaka <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Fri, Apr 20, 2018 at 11:40 AM, Kyotaro HORIGUCHI
<[email protected]> wrote:
> By the way, I think I found a bug of FPW.
>
> The following steps yields INSERT record that doesn't have a FPI
> after a checkpoint.
>
> (Start server with full_page_writes = off)
> CREATE TABLE t (a int);
> CHECKPOINT;
> INSERT INTO t VALUES (1);
> ALTER SYSTEM SET full_page_writes TO on;
> SELECT pg_reload_conf();
> CHECKPOINT;
> INSERT INTO t VALUES (1);
>
> The last insert is expected to write a record with FPI but it
> doesn't actually. No FPI will be written for the page after that.
>
> It seems that the reason is that XLogInsertRecord is forgetting
> to check doPageWrites' update.
>
> In the failure case, fpw_lsn has been set by XLogRecordAssemble
> but doPageWrite is false at the time and it considers that no FPI
> is required then it sets fpw_lsn to InvalidXLogRecPtr.
>
> After that, XLogInsertRecord receives the record but it thinks
> that doPageWrites is true since it looks the shared value.
>
>> if (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr && doPageWrites)
>
> So this line thinks that "no FPI is omitted in this record" but
> actually the record is just forgotten to attach them.
>
> The attached patch lets XLogInsertRecord check if doPageWrites
> has been turned on after the last call and cause recomputation in
> the case.
>

I think you have correctly spotted the problem and your fix looks good
to me.  As this is a separate problem and fix is different from what
we are discussing here, I think this can be committed it separately.

-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-23 16:21  Robert Haas <[email protected]>
  parent: Michael Paquier <[email protected]>
  2 siblings, 1 reply; 292+ messages in thread

From: Robert Haas @ 2018-04-23 16:21 UTC (permalink / raw)
  To: Michael Paquier <[email protected]>; +Cc: Amit Kapila <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Wed, Apr 18, 2018 at 9:49 PM, Michael Paquier <[email protected]> wrote:
> On Wed, Apr 18, 2018 at 10:52:51AM -0400, Robert Haas wrote:
>> I would just document the risks.  If the documentation says that you
>> can't rely on the value until after the next checkpoint, or whatever
>> the rule is, then I think we're fine.  I don't think that we really
>> have the infrastructure to do any better; if we try, we'll just end up
>> with odd warts.  Documenting the current set of warts is less churn
>> and less work.
>
> The last version of the patch proposed has eaten this diff which was
> part of one of the past versions (v2-0001-Change-FPW-handling.patch from
> https://www.postgresql.org/message-id/20180412.103430.133595350.horiguchi.kyotaro%40lab.ntt.co.jp):
> +        The default is <literal>on</literal>. The change of the parameter takes
> +        effect at the next checkpoint time.
> So there were some documentation about the beHavior change for what it's
> worth.

Fine, but that doesn't answer the question of whether we actually need
to or should change the behavior in the first place.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company




^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-23 23:52  Michael Paquier <[email protected]>
  parent: Robert Haas <[email protected]>
  0 siblings, 1 reply; 292+ messages in thread

From: Michael Paquier @ 2018-04-23 23:52 UTC (permalink / raw)
  To: Robert Haas <[email protected]>; +Cc: Amit Kapila <[email protected]>; Heikki Linnakangas <[email protected]>; Kyotaro HORIGUCHI <[email protected]>; Dilip Kumar <[email protected]>; pgsql-hackers

On Mon, Apr 23, 2018 at 12:21:04PM -0400, Robert Haas wrote:
> Fine, but that doesn't answer the question of whether we actually need
> to or should change the behavior in the first place.

Per the last arguments that would be "No, we don't want to change it as
it would surprise some users":
https://www.postgresql.org/message-id/[email protected]
--
Michael


Attachments:

  [application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
  download

^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* Re: Problem while setting the fpw with SIGHUP
@ 2018-04-24 01:40  Kyotaro HORIGUCHI <[email protected]>
  parent: Michael Paquier <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Kyotaro HORIGUCHI @ 2018-04-24 01:40 UTC (permalink / raw)
  To: [email protected]; +Cc: [email protected]; [email protected]; [email protected]; [email protected]; pgsql-hackers

At Tue, 24 Apr 2018 08:52:17 +0900, Michael Paquier <[email protected]> wrote in <[email protected]>
> On Mon, Apr 23, 2018 at 12:21:04PM -0400, Robert Haas wrote:
> > Fine, but that doesn't answer the question of whether we actually need
> > to or should change the behavior in the first place.
> 
> Per the last arguments that would be "No, we don't want to change it as
> it would surprise some users":
> https://www.postgresql.org/message-id/[email protected]

The answer is that the change of behavior is not required to fix
the bug. So I'm fine with applying only (0001 and) 0002 here.

https://www.postgresql.org/message-id/[email protected]

One reason that I introduced the "restriction" was that it was
useful to avoid concurrent udpate of the flag. It is now avoided
in another way.

Just my opinion on the behavior follows.

I don't think anyone confirms that FPI come back after switching
full_page_writes (one of the reason is it needs pg_waldump).
pg_start/stop_backup() on standby says that "You need to turn on
FPW, then do checkpoint". It suggests that FPI's don't work
before the next checkpoint. If we keep the current behavior, the
documentation might need an additional phrase something like "FPW
ensures that data is protected from partial writes after the next
chackpoint starts". On the other hand honestly I don't have
confidence that the WAL reduction worth the additional complexity
by 0003.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread

* [PATCH v6] README for 64bit xid
@ 2022-01-10 19:20  Pavel Borisov <[email protected]>
  0 siblings, 0 replies; 292+ messages in thread

From: Pavel Borisov @ 2022-01-10 19:20 UTC (permalink / raw)

Authors:
- Pavel Borisov <[email protected]>
- Maxim Orlov <[email protected]>
- Yura Sokolov <[email protected]> <[email protected]>
---
 src/backend/access/heap/README.XID64 | 128 +++++++++++++++++++++++++++
 1 file changed, 128 insertions(+)
 create mode 100644 src/backend/access/heap/README.XID64

diff --git a/src/backend/access/heap/README.XID64 b/src/backend/access/heap/README.XID64
new file mode 100644
index 00000000000..457ba9b9ef5
--- /dev/null
+++ b/src/backend/access/heap/README.XID64
@@ -0,0 +1,128 @@
+src/backend/access/heap/README.XID64
+
+64-bit Transaction ID's (XID)
+=============================
+
+A limited number (N = 2^32) of XID's required to do vacuum freeze to prevent
+wraparound every N/2 transactions. This causes performance degradation due
+to the need to exclusively lock tables while being vacuumed. In each
+wraparound cycle, SLRU buffers are also being cut.
+
+With 64-bit XID's wraparound is effectively postponed to a very distant
+future. Even in highly loaded systems that had 2^32 transactions per day
+it will take huge 2^31 days before the first enforced "vacuum to prevent
+wraparound"). Buffers cutting and routine vacuum are not enforced, and DBA
+can plan them independently at the time with the least system load and least
+critical for database performance. Also, it can be done less frequently
+(several times a year vs every several days) on systems with transaction rates
+similar to those mentioned above.
+
+On-disk tuple and page format
+-----------------------------
+
+On-disk tuple format remains unchanged. 32-bit t_xmin and t_xmax store the
+lower parts of 64-bit XMIN and XMAX values. Each heap page has additional
+64-bit pd_xid_base and pd_multi_base which are common for all tuples on a page.
+They are placed into a pd_special area - 16 bytes in the end of a heap page.
+Actual XMIN/XMAX for a tuple are calculated upon reading a tuple from a page
+as follows:
+
+XMIN = t_xmin + pd_xid_base. 					(1)
+XMAX = t_xmax + pd_xid_base/pd_multi_base.		(2)
+
+"Double XMAX" page format
+---------------------------------
+
+At first read of a heap page after pg_upgrade from 32-bit XID PostgreSQL
+version pd_special area with a size of 16 bytes should be added to a page.
+Though a page may not have space for this. Then it can be converted to a
+temporary format called "double XMAX".
+
+All tuples after pg-upgrade would necessarily have xmin = FrozenTransactionId.
+So we don't need tuple header t_xmin field and we reuse t_xmin to store higher
+32 bits of its XMAX.
+
+Double XMAX format is only for full pages that don't have 16 bytes for
+pd_special. So it neither has a�place for a single tuple. Insert and HOT update
+for double XMAX pages is impossible and not supported. We can only read or
+delete tuples from it.
+
+When we are able to prune page double XMAX it will be converted from it to
+general 64-bit XID page format with all operations on its tuples supported.
+
+In-memory tuple format
+----------------------
+
+In-memory tuple representation consists of two parts:
+- HeapTupleHeader from disk page (contains all heap tuple contents, not only
+header)
+- HeapTuple with additional in-memory fields
+
+HeapTuple for each tuple in memory stores t_xid_base/t_multi_base - a copies of
+page's pd_xid_base/pd_multi_base. With tuple's 32-bit t_xmin and t_xmax from
+HeapTupleHeader they are used to calculate actual 64-bit XMIN and XMAX:
+
+XMIN = t_xmin + t_xid_base. 					(3)
+XMAX = t_xmax + t_xid_base/t_multi_base.		(4)
+
+The downside of this is that we can not use tuple's XMIN and XMAX right away.
+We often need to re-read t_xmin and t_xmax - which could actually be pointers
+into a page in shared buffers and therefore they could be updated by any other
+backend.
+
+Update/delete with 64-bit XIDs and 32-bit t_xmin/t_xmax
+--------------------------------------------------------------
+
+When we try to delete/update a tuple, we check that XMAX for a page fits (2).
+I.e. that t_xmax will not be over MaxShortTransactionId relative to
+pd_xid_base/pd_multi_base of a its page.
+
+If the current XID doesn't fit a range
+(pd_xid_base, pd_xid_base + MaxShortTransactionId) (5):
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base/pd_multi_base on
+a page and update all t_xmin/t_xmax of the other tuples on the page to
+correspond new pd_xid_base/pd_multi_base.
+
+- If it was impossible, it will try to prune and freeze tuples on a page.
+
+- If this is unsuccessful it will throw an error. Normally this is very
+unlikely but if there is a very old living transaction with an age of around
+2^32 this can arise. Basically, this is a behavior similar to one during the
+vacuum to prevent wraparound when XID was 32-bit. Dba should take care and
+avoid very-long-living transactions with an age close to 2^32. So long-living
+transactions often they are most likely defunct.
+
+Insert with 64-bit XIDs and 32-bit t_xmin/t_xmax
+------------------------------------------------
+
+On insert we check if current XID fits a range (5). Otherwise:
+
+- heap_page_prepare_for_xid() will try to increase pd_xid_base for t_xmin will
+not be over MaxShortTransactionId.
+
+- If it is impossible, then it will try to prune and freeze tuples on a page.
+
+Known issue: if pd_xid_base could not be shifted to accommodate a tuple being
+inserted due to a very long-running transaction, we just throw an error. We
+neither try to insert a�tuple into another page nor mark the current page as
+full. So, in this (unlikely) case we will get regular insert errors on the next
+tries to insert to the page 'locked' by this very long-running transaction.
+
+Upgrade from 32-bit XID versions
+--------------------------------
+
+pg_upgrade doesn't change pages format itself. It is done lazily after.
+
+1. At first heap page read, tuples on a page are repacked to free 16 bytes
+at the end of a page, possibly freeing space from dead tuples.
+
+2A. 16 bytes of pd_special is added if there is a place for it
+
+2B. Page is converted to "Double XMAX" format if there is no place for
+pd_special
+
+3. If a page is in double XMAX format after its first read, and vacuum (or
+micro-vacuum at select query) could prune some tuples and free space for
+pd_special, prune_page will add pd_special and convert page from double XMAX
+to general 64-bit XID page format.
-- 
2.24.3 (Apple Git-128)


--cpok4wp6gsarlzvp--





^ permalink  raw  reply  [nested|flat] 292+ messages in thread


end of thread, other threads:[~2022-01-10 19:20 UTC | newest]

Thread overview: 292+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2018-04-04 08:26 Re: Problem while setting the fpw with SIGHUP Kyotaro HORIGUCHI <[email protected]>
2018-04-06 08:20 ` Kyotaro HORIGUCHI <[email protected]>
2018-04-06 12:29   ` Amit Kapila <[email protected]>
2018-04-09 08:13     ` Kyotaro HORIGUCHI <[email protected]>
2018-04-11 11:09       ` Heikki Linnakangas <[email protected]>
2018-04-11 20:24         ` Michael Paquier <[email protected]>
2018-04-12 01:34           ` Kyotaro HORIGUCHI <[email protected]>
2018-04-12 05:07             ` Michael Paquier <[email protected]>
2018-04-12 07:59               ` Kyotaro HORIGUCHI <[email protected]>
2018-04-12 08:10                 ` Michael Paquier <[email protected]>
2018-04-12 18:55         ` Robert Haas <[email protected]>
2018-04-13 01:29           ` Michael Paquier <[email protected]>
2018-04-13 03:01             ` Amit Kapila <[email protected]>
2018-04-13 04:47               ` Kyotaro HORIGUCHI <[email protected]>
2018-04-13 08:28                 ` Kyotaro HORIGUCHI <[email protected]>
2018-04-13 08:34                   ` Kyotaro HORIGUCHI <[email protected]>
2018-04-13 17:06             ` Robert Haas <[email protected]>
2018-04-18 14:37               ` Amit Kapila <[email protected]>
2018-04-18 14:52                 ` Robert Haas <[email protected]>
2018-04-19 01:49                   ` Michael Paquier <[email protected]>
2018-04-19 13:41                     ` Amit Kapila <[email protected]>
2018-04-20 01:04                       ` Michael Paquier <[email protected]>
2018-04-20 08:33                       ` Kyotaro HORIGUCHI <[email protected]>
2018-04-20 06:10                     ` Kyotaro HORIGUCHI <[email protected]>
2018-04-20 12:36                       ` Amit Kapila <[email protected]>
2018-04-23 16:21                     ` Robert Haas <[email protected]>
2018-04-23 23:52                       ` Michael Paquier <[email protected]>
2018-04-24 01:40                         ` Kyotaro HORIGUCHI <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>
2022-01-10 19:20 [PATCH v6] README for 64bit xid Pavel Borisov <[email protected]>

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox