MIME-Version: 1.0
References: <f15d665f-4cd1-4894-037c-afdbe369287e@gmail.com>
In-Reply-To: <f15d665f-4cd1-4894-037c-afdbe369287e@gmail.com>
From: vignesh C <vignesh21@gmail.com>
Date: Wed, 29 May 2024 21:26:12 +0530
Message-ID: <CALDaNm0weOgCmcST0mLBPxWdPm=BVeZ35GMxTWwPbSX6bjcRWw@mail.gmail.com>
Subject: Re: 001_rep_changes.pl fails due to publisher stuck on shutdown
To: Alexander Lakhin <exclusion@gmail.com>
Cc: pgsql-hackers <pgsql-hackers@postgresql.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: <https://www.postgresql.org/message-id/CALDaNm0weOgCmcST0mLBPxWdPm%3DBVeZ35GMxTWwPbSX6bjcRWw%40mail.gmail.com>
Precedence: bulk

On Wed, 29 May 2024 at 16:30, Alexander Lakhin <exclusion@gmail.com> wrote:
>
> Hello hackers,
>
> As a recent buildfarm test failure [1] shows:
> [14:33:02.374](0.333s) ok 23 - update works with dropped subscriber colum=
n
> ### Stopping node "publisher" using mode fast
> # Running: pg_ctl -D
> /home/bf/bf-build/adder/HEAD/pgsql.build/testrun/subscription/001_rep_cha=
nges/data/t_001_rep_changes_publisher_data/pgdata
> -m fast stop
> waiting for server to shut down.. ... ... ... .. failed
> pg_ctl: server does not shut down
> # pg_ctl stop failed: 256
> # Postmaster PID for node "publisher" is 2222549
> [14:39:04.375](362.001s) Bail out!  pg_ctl stop failed
>
> 001_rep_changes_publisher.log
> 2024-05-16 14:33:02.907 UTC [2238704][client backend][4/22:0] LOG: statem=
ent: DELETE FROM tab_rep
> 2024-05-16 14:33:02.925 UTC [2238704][client backend][:0] LOG: disconnect=
ion: session time: 0:00:00.078 user=3Dbf
> database=3Dpostgres host=3D[local]
> 2024-05-16 14:33:02.939 UTC [2222549][postmaster][:0] LOG:  received fast=
 shutdown request
> 2024-05-16 14:33:03.000 UTC [2222549][postmaster][:0] LOG:  aborting any =
active transactions
> 2024-05-16 14:33:03.049 UTC [2222549][postmaster][:0] LOG: background wor=
ker "logical replication launcher" (PID
> 2223110) exited with exit code 1
> 2024-05-16 14:33:03.062 UTC [2222901][checkpointer][:0] LOG: shutting dow=
n
> 2024-05-16 14:39:04.377 UTC [2222549][postmaster][:0] LOG:  received imme=
diate shutdown request
> 2024-05-16 14:39:04.382 UTC [2222549][postmaster][:0] LOG:  database syst=
em is shut down
>
> the publisher node may hang on stopping.
>
> I reproduced the failure (with aggressive autovacuum) locally and
> discovered that it happens because:
> 1) checkpointer calls WalSndInitStopping() (which sends
>   PROCSIG_WALSND_INIT_STOPPING to walsender), and then spins inside
>   WalSndWaitStopping() indefinitely, because:
> 2) walsender receives the signal, sets got_STOPPING =3D true, but can't e=
xit
> WalSndLoop():
> 3) it never sets got_SIGUSR2 (to get into WalSndDone()) in
>   XLogSendLogical():
> 4) it never sets WalSndCaughtUp to true:
> 5) logical_decoding_ctx->reader->EndRecPtr can't reach flushPtr in
>   XLogSendLogical():
> 6) EndRecPtr doesn't advance in XLogNextRecord():
> 7) XLogDecodeNextRecord() fails do decode a record that crosses a page
>   boundary:
> 8) ReadPageInternal() (commented "Wait for the next page to become
>   available") constantly returns XLREAD_FAIL:
> 9) state->routine.page_read()/logical_read_xlog_page() constantly returns
>   -1:
> 10) flushptr =3D WalSndWaitForWal() stops advancing, because
>   got_STOPPING =3D=3D true (see 2).
>
> That is, walsender doesn't let itself to catch up, if it gets the stop
> signal when it's lagging behind and decoding a record requires reading
> the next wal page.
>
> Please look at the reproducing test (based on 001_rep_changes.pl) attache=
d.
> If fails for me as below:
> # 17
> Bailout called.  Further testing stopped:  pg_ctl stop failed
> FAILED--Further testing stopped: pg_ctl stop failed
> make: *** [Makefile:21: check] =D0=9E=D1=88=D0=B8=D0=B1=D0=BA=D0=B0 255

Thank you, Alexander, for sharing the script. I was able to reproduce
the issue using the provided script. Furthermore, while investigating
its origins, I discovered that this problem persists across all
branches up to PG10 (the script needs slight adjustments to run it on
older versions). It's worth noting that this issue isn't a result of
recent version changes.

Regards,
Vignesh