Feedback-ID: i0fe9450f:Fastmail
Date: Thu, 7 Sep 2023 16:25:27 +0900
From: Michael Paquier <michael@paquier.xyz>
To: Amit Kapila <amit.kapila16@gmail.com>
Cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>,
	Sawada Masahiko <sawada.mshk@gmail.com>,
	Dilip Kumar <dilipbalaut@gmail.com>,
	"Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>,
	vignesh C <vignesh21@gmail.com>,
	=?utf-8?B?SG91LCBaaGlqaWUv5L6vIOW/l+adsA==?= <houzj.fnst@fujitsu.com>,
	Peter Smith <smithpb2250@gmail.com>
Subject: Re: Impact of checkpointer during pg_upgrade
Message-ID: <ZPl659a5hPDHPq9w@paquier.xyz>
References: <CAA4eK1LV3+76CSOAk0h8Kv0AKb-OETsJHe6Sq6172-7DZXf0Qg@mail.gmail.com>
MIME-Version: 1.0
Content-Type: multipart/signed; micalg=pgp-sha512;
	protocol="application/pgp-signature"; boundary="zVL4WyyJzzoQg5T1"
Content-Disposition: inline
In-Reply-To: <CAA4eK1LV3+76CSOAk0h8Kv0AKb-OETsJHe6Sq6172-7DZXf0Qg@mail.gmail.com>
Archived-At: <https://www.postgresql.org/message-id/ZPl659a5hPDHPq9w%40paquier.xyz>
Precedence: bulk


--zVL4WyyJzzoQg5T1
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

(Just dumping what I have in mind while reading the thread.)

On Sat, Sep 02, 2023 at 10:08:51AM +0530, Amit Kapila wrote:
> During pg_upgrade, we start the server for the old cluster which can
> allow the checkpointer to remove the WAL files. It has been noticed
> that we do generate certain types of WAL records (e.g
> XLOG_RUNNING_XACTS, XLOG_CHECKPOINT_ONLINE, and XLOG_FPI_FOR_HINT)
> even during pg_upgrade for old cluster, so additional WAL records
> could let checkpointer decide that certain WAL segments can be removed
> (e.g. say wal size crosses max_slot_wal_keep_size_mb) and invalidate
> the slots. Currently, I can't see any problem with this but for future
> work where we want to migrate logical slots during an upgrade[1], we
> need to decide what to do for such cases. The initial idea we had was
> that if the old cluster has some invalid slots, we won't allow an
> upgrade unless the user removes such slots or uses some option like
> --exclude-slots. It is quite possible that slots got invalidated
> during pg_upgrade due to no user activity. Now, even though the
> possibility of the same is less I think it is worth considering what
> should be the behavior.
>=20
> The other possibilities apart from not allowing an upgrade in such a
> case could be (a) Before starting the old cluster, we fetch the slots
> directly from the disk using some tool like [2] and make the decisions
> based on that state; (b) During the upgrade, we don't allow WAL to be
> removed if it can invalidate slots; (c) Copy/Migrate the invalid slots
> as well but for that, we need to expose an API to invalidate the
> slots; (d) somehow distinguish the slots that are invalidated during
> an upgrade and then simply copy such slots because anyway we ensure
> that all the WAL required by slot is sent before shutdown.

Checking for any invalid slots at the beginning of the upgrade and
complain sounds like a good thing to do, because these are not
expected to begin with, no?  That's similar to a pre-check requirement
that the slots should have fed the subscribers with all the data
available up to the shutdown checkpoint when the publisher was stopped
before running pg_upgrade.  So (a) is a good idea to prevent an
upgrade based on a state we don't expect from the start, as something
in check.c, I assume.

On top of that, (b) sounds like a good idea to me anyway to be more
defensive.  But couldn't we do that just like we do for autovacuum and
force the GUCs that could remove the slot's WAL to not do their work
instead?  An upgrade is a special state of the cluster, and I'm not
much into painting more checks based on IsBinaryUpgrade to prevent WAL
segments to be removed while we can have a full control of what we
want with just the GUCs that force the hand of the slots.  That just
seems simpler to me, and the WAL recycling logic is already complex
particularly with all the GUCs popping lately to force some conditions
to do the WAL recycling and/or removal.
=20
During the upgrade, if some segments are removed and some of the slots
we expect to still be valid once the upgrade is done are marked as
invalid and become unusable, I think that we should just copy these
slots, but also update their state data so as they can still be used
with what we expect, as these could be wanted by the subscribers.
That's what you mean with (d), I assume.  Do you think that it would
be possible to force the slot's data on the publishers so as they use
a local LSN based on the new WAL we are resetting at?  At least that
seems more friendly to me as this limits the set of manipulations to
do on the slots for the end-user.  The protection from (b) ought to be
enough, in itself.  (c) overlaps with (a), especially if we want to be
able to read or write some of the slot's data offline.
--
Michael

--zVL4WyyJzzoQg5T1
Content-Type: application/pgp-signature; name="signature.asc"

-----BEGIN PGP SIGNATURE-----
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=8QJQ
-----END PGP SIGNATURE-----

--zVL4WyyJzzoQg5T1--