Re: SIGSEGV from START_REPLICATION 0/XXXXXXX in XLogSendPhysical () at walsender.c:2762

Kyotaro Horiguchi <horikyota.ntt@gmail.com>

From: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
To: michael@paquier.xyz
Cc: sitnikov.vladimir@gmail.com, pgsql-hackers@postgresql.org
Date: 2020-05-29T08:56:53Z
Lists: pgsql-hackers
At Fri, 29 May 2020 16:21:38 +0900, Michael Paquier <michael@paquier.xyz> wrote in 
> On Thu, May 28, 2020 at 06:11:39PM +0900, Kyotaro Horiguchi wrote:
> > Mmm. It is not the proper way to use physical replication and it's
> > totally accidental that that worked (or even it might be a bug). The
> > documentation is saying as the follows, as more-or-less the same for
> > all versions since 9.4.
> > 
> > https://www.postgresql.org/docs/13/protocol-replication.html
> 
> +       if (am_db_walsender)
> +               ereport(ERROR,
> +                       (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
> +                                errmsg("cannot initiate physical
> replication on a logical replication connection")));
> 
> I don't agree with this change.  The only restriction that we have in
> place now in walsender.c regarding MyDatabaseId not being set is to
> prevent the execution of SQL commands.  Note that it is possible to
> start physical replication even if MyDatabaseId is set in a
> replication connection, so you could break cases that have been valid
> until now.

It donesn't check MyDatabase, but whether the connection parameter
"repliation" is "true" or "database".  The documentation is telling
that "replication" should be "true" for a connection that is to be
used for physical replication, and "replication" should literally be
"database" for a connection that is for logical replication.  We need
to revise the documentation if we are going to allow physical
replication on a conection with "replication = database".

> I think that we actually should be much more careful with the
> initialization of the WAL reader used in the context of a WAL sender
> before calling WALRead() and attempting to read a new WAL page.

I agree that the initialization can be improved, but the current code
is no problem if we don't allow to run both logical and physical
replication on a single session.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center



Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →
  1. Fix crash in WAL sender when starting physical replication

  2. Add a SHOW command to the replication command language.