Date: Mon, 24 Jun 2024 15:08:45 -0700
From: Noah Misch <noah@leadboat.com>
To: Amit Kapila <amit.kapila16@gmail.com>
Cc: Peter Eisentraut <peter@eisentraut.org>,
	Euler Taveira <euler@eulerto.com>,
	Shlok Kyal <shlok.kyal.oss@gmail.com>,
	"kuroda.hayato@fujitsu.com" <kuroda.hayato@fujitsu.com>,
	Tomas Vondra <tomas.vondra@enterprisedb.com>,
	"pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>,
	Michael Paquier <michael@paquier.xyz>,
	Andres Freund <andres@anarazel.de>,
	Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>,
	=?iso-8859-1?Q?Fabr=EDzio?= de Royes Mello <fabriziomello@gmail.com>,
	vignesh C <vignesh21@gmail.com>
Subject: Re: speed up a logical replica setup
Message-ID: <20240624220845.ba.nmisch@google.com>
References: <2f0b73f3-e7c0-4e95-9712-1993442a6b64@app.fastmail.com>
 <CAA4eK1JtnzqYNQ_S+mj_XsPfgk7szxegKiVQq4-8fY1wUcUnpg@mail.gmail.com>
 <34637e7f-0330-420d-8f45-1d022962d2fe@app.fastmail.com>
 <7a970912-0b77-4942-84f7-2c9ca0bc05a5@eisentraut.org>
 <d546c4bb-92d1-4e2d-898f-48234b12ed25@app.fastmail.com>
 <CANhcyEWdTTO71v+GFYUTsOOUvHus1qccg88bRmsU9KykJqjSrQ@mail.gmail.com>
 <44de4e04-db22-40ed-8ba9-46387dc8bf83@app.fastmail.com>
 <3fa9ef0f-b277-4c13-850a-8ccc04de1406@eisentraut.org>
 <20240623062157.97.nmisch@google.com>
 <CAA4eK1JGyqFXfvXD+Z81WNz=gOSq1_Ze8MqdLE9PMvTWnVpuZA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <CAA4eK1JGyqFXfvXD+Z81WNz=gOSq1_Ze8MqdLE9PMvTWnVpuZA@mail.gmail.com>
User-Agent: Mutt/2.2.12 (2023-09-09)
Archived-At: <https://www.postgresql.org/message-id/20240624220845.ba.nmisch%40google.com>
Precedence: bulk

On Mon, Jun 24, 2024 at 05:20:21PM +0530, Amit Kapila wrote:
> On Sun, Jun 23, 2024 at 11:52 AM Noah Misch <noah@leadboat.com> wrote:
> >
> > > +static void
> > > +create_publication(PGconn *conn, struct LogicalRepInfo *dbinfo)
> > > +{
> >
> > > +     appendPQExpBuffer(str, "CREATE PUBLICATION %s FOR ALL TABLES",
> > > +                                       ipubname_esc);
> >
> > This tool's documentation says it "guarantees that no transaction will be
> > lost."  I tried to determine whether achieving that will require something
> > like the fix from
> > https://postgr.es/m/flat/de52b282-1166-1180-45a2-8d8917ca74c6@enterprisedb.com.
> > (Not exactly the fix from that thread, since that thread has not discussed the
> > FOR ALL TABLES version of its race condition.)  I don't know.  On the one
> > hand, pg_createsubscriber benefits from creating a logical slot after creating
> > the publication.  That snapbuild.c process will wait for running XIDs.  On the
> > other hand, an INSERT/UPDATE/DELETE acquires its RowExclusiveLock and builds
> > its relcache entry before assigning an XID, so perhaps the snapbuild.c process

Correction: it doesn't matter how the original INSERT/UPDATE/DELETE builds its
relcache entry, just how pgoutput of the change builds the relcache entry from
the historic snapshot.

> > isn't enough to prevent that thread's race condition.  What do you think?
> 
> I am not able to imagine how the race condition discussed in the
> thread you quoted can impact this patch. The problem discussed is
> mainly the interaction when we are processing the changes in logical
> decoding w.r.t concurrent DDL (Alter Publication ... Add Table). The
> problem happens because we use the old cache state.

Right.  Taking the example from
http://postgr.es/m/20231119021830.d6t6aaxtrkpn743y@awork3.anarazel.de, LSNs
between what that mail calls 4) and 5) are not safely usable as start points.
pg_createsubscriber evades that thread's problem if the consistent_lsn it
passes to pg_replication_origin_advance() can't be in a bad-start-point LSN
span.  I cautiously bet the snapbuild.c process achieves that:

> I am missing your
> point about the race condition mentioned in the thread you quoted with
> snapbuild.c. Can you please elaborate a bit more?

When pg_createsubscriber calls pg_create_logical_replication_slot(), the key
part starts at:

	/*
	 * If caller needs us to determine the decoding start point, do so now.
	 * This might take a while.
	 */
	if (find_startpoint)
		DecodingContextFindStartpoint(ctx);

Two factors protect pg_createsubscriber.  First, (a) CREATE PUBLICATION
committed before pg_create_logical_replication_slot() started.  Second, (b)
DecodingContextFindStartpoint() waits for running XIDs to complete, via the
process described at the snapbuild.c "starting up in several stages" diagram.
Hence, the consistent_lsn is not in a bad-start-point LSN span.  It's fine
even if the original INSERT populated all caches before CREATE PUBLICATION
started and managed to assign an XID only after consistent_lsn.  From the
pgoutput perspective, that's indistinguishable from the transaction starting
at its first WAL record, after consistent_lsn.  The linked "long-standing data
loss bug in initial sync of logical replication" thread doesn't have (a),
hence its bug.  How close is that to accurate?

Thanks,
nm