Horton

Re: Apparent walsender bug triggered by logical replication

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: pgsql-hackers@postgreSQL.org
Date: 2017-06-30T00:15:33Z
Lists: pgsql-hackers 
Hi,

On 2017-06-29 20:07:14 -0400, Tom Lane wrote:
> I was able to make the hang go away by means of the attached patch that
> allows WalSndWaitForWal to exit early if the client has shut down the
> COPY.  However, since that function is miserably underdocumented (like
> most of this code :-(), I have little idea if this is correct or safe.

Seems reasonable to me.


> I also wonder why WalSndWaitForWal is being called for WAL that seemingly
> doesn't exist yet, and whether that doesn't indicate another bug somewhere
> in this stack.

That's pretty normal - we can only send back something once a
transaction is complete, and until that happens we'll just block waiting
for more WAL.

Greetings,

Andres Freund

Commits

  1. Reduce delay for last logicalrep feedback message when master goes idle.

  2. Shorten timeouts while waiting for logicalrep worker slot attach/detach.

  3. Fix walsender to exit promptly if client requests shutdown.