Content-Type: text/plain;
	charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 16.0 \(3864.400.21\))
Subject: Re: eliminate xl_heap_visible to reduce WAL (and eventually set VM
 on-access)
From: Chao Li <li.evan.chao@gmail.com>
In-Reply-To: 
 <CAAKRu_a7ByG2LipFADR7UYnLP4BWQKOV1sajqJC4=R37iO05+A@mail.gmail.com>
Date: Wed, 4 Mar 2026 16:59:41 +0800
Cc: Andres Freund <andres@anarazel.de>,
 Andrey Borodin <x4mmm@yandex-team.ru>,
 Kirill Reshke <reshkekirill@gmail.com>,
 Xuneng Zhou <xunengzhou@gmail.com>,
 Robert Haas <robertmhaas@gmail.com>,
 PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>,
 Heikki Linnakangas <hlinnaka@iki.fi>
Content-Transfer-Encoding: quoted-printable
Message-Id: <C3AB3F5B-626E-4AAA-9529-23E9A20C727F@gmail.com>
References: <6BC5DBAB-6084-4BB8-8450-52E9648AB021@gmail.com>
 <CAAKRu_ZCjHoRPfQ8AbMrFY8TOMCPAvZ0_m9SX7yg0edfTk45-g@mail.gmail.com>
 <7F5BCD7A-764D-4D8D-8E27-6F2CAAEA1CEE@gmail.com>
 <CAAKRu_a04jbDACwzRYwzDND31aPyf7Yvz9TAZrTr=+F5bK1aVA@mail.gmail.com>
 <CALdSSPjcv25jmXm29X-MRWZBae6+HwcWfVH1PE8NfD=EMTnkAg@mail.gmail.com>
 <CAAKRu_bwtBEzDwemyim1r6yYonw7FTyFr1HXG8vywCe-MdbPBQ@mail.gmail.com>
 <4379FDA3-9446-4E2C-9C15-32EFE8D4F31B@yandex-team.ru>
 <CAAKRu_YQd=2KvomM+RHcpeDKj0bq+peJ=3W-fip+pkvzA-Jq9w@mail.gmail.com>
 <7ib3sa55sapwjlaz4sijbiq7iezna27kjvvvar4dpgkmadml6t@gfpkkwmdnepx>
 <CAAKRu_bs+gZ83QDacmBxunPvCGnXJ05hxP2BDPJ3BGwdbGRXzg@mail.gmail.com>
 <bqc4kh5midfn44gnjiqez3bjqv4zogydguvdn446riw45jcf3y@4ez66il7ebvk>
 <CAAKRu_Y1MuANdm1p47Ev13Y9EQz8z+pw-vHOh=3DVdahUTjgXg@mail.gmail.com>
 <BD8B69E7-26D8-4706-9164-597C6AE57812@gmail.com>
 <CAAKRu_a7ByG2LipFADR7UYnLP4BWQKOV1sajqJC4=R37iO05+A@mail.gmail.com>
To: Melanie Plageman <melanieplageman@gmail.com>
Archived-At: 
 <https://www.postgresql.org/message-id/C3AB3F5B-626E-4AAA-9529-23E9A20C727F%40gmail.com>
Precedence: bulk


> On Mar 3, 2026, at 23:52, Melanie Plageman <melanieplageman@gmail.com> =
wrote:
>=20
>=20
>> Otherwise, if prstate->pagefrz.FreezePageConflictXid is still =
possibly be InvalidTransactionId, then the Assert should be changed to =
something like:
>>=20
>> Assert(prstate->pagefrz.FreezePageConflictXid =3D=3D =
InvalidTransactionId ||
>>  =
TransactionIdPrecedesOrEquals(prstate->pagefrz.FreezePageConflictXid, =
prstate->cutoffs->OldestXmin)
>=20
> This is covered by TransactionIdPrecedesOrEquals because
> InvalidTransactionId is 0. We assume that in many places throughout
> the code.
>=20

I understood that TransactionIdPrecedesOrEquals(InvalidTransactionId, =
prstate->cutoffs->OldestXmin) is true, but that would leave an =
impression to code readers that prstate->pagefrz.FreezePageConflictXid =
could not be InvalidTransactionId. Thus I think my version explicitly =
tells that prstate->pagefrz.FreezePageConflictXid could be =
InvalidTransactionId at the point.


>> I will continue with 0005 tomorrow.
>=20

4 - 0005
```
  * Caller must have pin on the buffer, and must *not* have a lock on =
it.
  */
 void
-heap_page_prune_opt(Relation relation, Buffer buffer)
+heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer)
```

I don=E2=80=99t see why vmbuffer has to be of pointer type. Buffer type =
is underlying int, I checked the last commit, vmbuffer only passes in =
data into the function without passing out anything.

As we add the new parameter vmbuffer, though it=E2=80=99s not used in =
this commit, I think it=E2=80=99d be better to update the header commit =
to explain what this parameter will do.

5  - 0006
```
+ *
+ * heap_fix_vm_corruption() makes changes to the VM and, potentially, =
the heap
+ * page, but it does not need to be done in a critical section because
+ * clearing the VM is not WAL-logged.
+ */
+static void
+heap_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum)
```

Nit: why the last paragraph of the header comments uses the function =
name instead of =E2=80=9Cthis function=E2=80=9D? Looks like a =
copy-pasto.

6 - 0006
```
+		if (prstate->lpdead_items > 0)
+		{
+			ereport(WARNING,
+					=
(errcode(ERRCODE_DATA_CORRUPTED),
+					 errmsg("LP_DEAD item found on =
page marked as all-visible"),
+					 errdetail("relation \"%s\", =
page %u, tuple %u",
+							   =
RelationGetRelationName(prstate->relation),
+							   =
prstate->block, offnum)));
+		}
+		else
+		{
+			ereport(WARNING,
+					=
(errcode(ERRCODE_DATA_CORRUPTED),
+					 errmsg("tuple not visible to =
all found on page marked as all-visible"),
+					 errdetail("relation \"%s\", =
page %u, tuple %u",
+							   =
RelationGetRelationName(prstate->relation),
+							   =
prstate->block, offnum)));
+		}
```

I recently just learned that a detail message should use complete =
sentences, and end each with a period, and capitalize the first word of =
sentences. See =
https://www.postgresql.org/docs/current/error-style-guide.html.

7 - 0006
```
+	else if (prstate->vmbits & VISIBILITYMAP_VALID_BITS)
+	{
+		/*
+		 * As of PostgreSQL 9.2, the visibility map bit should =
never be set if
+		 * the page-level bit is clear.  However, it's possible =
that the bit
+		 * got cleared after heap_vac_scan_next_block() was =
called, so we must
+		 * recheck with buffer lock before concluding that the =
VM is corrupt.
+		 */
+		ereport(WARNING,
+				(errcode(ERRCODE_DATA_CORRUPTED),
+				 errmsg("page %u in \"%s\" is not marked =
all-visible but visibility map bit is set",
+						prstate->block,
+						=
RelationGetRelationName(prstate->relation))));
+	}
```

The comment says =E2=80=9Cwe must recheck with buffer lock before=E2=80=A6=
=E2=80=9D, but it only log a warning message. Is the comment stale?

8 - 0007
```
+static void
+heap_page_bypass_prune_freeze(PruneState *prstate, PruneFreezeResult =
*presult)
+{
+	OffsetNumber maxoff =3D PageGetMaxOffsetNumber(prstate->page);
+	Page		page =3D prstate->page;
+
+	Assert(prstate->vmbits & VISIBILITYMAP_ALL_FROZEN ||
+		   (prstate->vmbits & VISIBILITYMAP_ALL_VISIBLE &&
+			!prstate->attempt_freeze));
+
+	/* We'll fill in presult for the caller */
+	memset(presult, 0, sizeof(PruneFreezeResult));
+
+	/*
+	 * Since the page is all-visible, a count of the normal ItemIds =
on the
+	 * page should be sufficient for vacuum's live tuple count.
+	 */
+	for (OffsetNumber off =3D FirstOffsetNumber;
+		 off <=3D maxoff;
+		 off =3D OffsetNumberNext(off))
+	{
+		if (ItemIdIsNormal(PageGetItemId(page, off)))
+			prstate->live_tuples++;
+	}
+
+	presult->live_tuples =3D prstate->live_tuples;
+
+	/* Clear any stale prune hint */
+	if (TransactionIdIsValid(PageGetPruneXid(page)))
+	{
+		PageClearPrunable(page);
+		MarkBufferDirtyHint(prstate->buffer, true);
+	}
+
+	presult->vmbits =3D prstate->vmbits;
+
+	if (!PageIsEmpty(page))
+		presult->hastup =3D true;
+}
```

* Given this function has done PageIsEmpty(page), that that is true, we =
don=E2=80=99t need to count live_tuples, right? That could be a tiny =
optimization.
* I see heap_page_bypass_prune_freeze() is only called in one place and =
immediately after prune_freeze_setup() and heap_fix_vm_corruption(), so =
prstate->vmbits must be 0, so do we need to do presult->vmbits =3D =
prstate->vmbits;?
* Do we need to set all_visible and all_frozen to presult?

0008 LGTM

I will continue with 0009 tomorrow.

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/