Hi,
On 2023-04-07 08:47:57 -0700, Andres Freund wrote:
> Integrated all of these.
Here's my current version. Changes:
- Integrated Bertrand's changes
- polished commit messages of 0001-0003
- edited code comments for 0003, including
InvalidateObsoleteReplicationSlots()'s header
- added a bump of SLOT_VERSION to 0001
- moved addition of pg_log_standby_snapshot() to 0007
- added a catversion bump for pg_log_standby_snapshot()
- moved all the bits dealing with procsignals from 0003 to 0004, now the split
makes sense IMO
- combined a few more sucessive ->safe_psql() calls
I see occasional failures in the tests, particularly in the new test using
pg_authid, but not solely. cfbot also seems to have seen these:
https://cirrus-ci.com/github/postgresql-cfbot/postgresql/commitfest%2F42%2F3740
I made a bogus attempt at a workaround for the pg_authid case last night. But
that didn't actually fix anything, it just changed the timing.
I think the issue is that VACUUM does not force WAL to be flushed at the end
(since it does not assign an xid). wait_for_replay_catchup() uses
$node->lsn('flush'), which, due to VACUUM not flushing, can be an LSN from
before VACUUM completed.
The problem can be made more likely by adding pg_usleep(1000000); before
walwriter.c's call to XLogBackgroundFlush().
We probably should introduce some infrastructure in Cluster.pm for this, but
for now I just added a 'flush_wal' table that we insert into after a
VACUUM. That guarantees a WAL flush.
I think some of the patches might have more reviewers than really applicable,
and might also miss some. I'd appreciate if you could go over that...
Greetings,
Andres Freund