Re: Race condition in recovery?

On Thu, May 27, 2021 at 2:26 AM Dilip Kumar <dilipbalaut(at)gmail(dot)com> wrote:
> Changed as suggested.

I don't think the code as written here is going to work on Windows,
because your code doesn't duplicate enable_restoring's call to
perl2host or its backslash-escaping logic. It would really be better
if we could use enable_restoring directly. Also, I discovered that the
'return' in cp_history_files should really say 'exit', because
otherwise it generates a complaint every time it's run. It should also
have 'use strict' and 'use warnings' at the top.

Here's a version of your test case patch with the 1-line code fix
added, the above issues addressed, and a bunch of cosmetic tweaks.
Unfortunately, it doesn't pass for me consistently. I'm not sure if
that's because I broke something with my changes, or because the test
contains an underlying race condition which we need to address.
Attached also are the log files from a failed run if you want to look
at them. The key lines seem to be:

2021-06-03 16:16:53.984 EDT [47796] LOG: restarted WAL streaming at
0/3000000 on timeline 2
2021-06-03 16:16:54.197 EDT [47813] 025_stuck_on_old_timeline.pl LOG:
statement: SELECT count(*) FROM tab_int
2021-06-03 16:16:54.197 EDT [47813] 025_stuck_on_old_timeline.pl
ERROR: relation "tab_int" does not exist at character 22

Or from the main log:

Waiting for replication conn cascade's replay_lsn to pass '0/3000000' on standby
done
error running SQL: 'psql:<stdin>:1: ERROR: relation "tab_int" does not exist
LINE 1: SELECT count(*) FROM tab_int
^'

I wonder whether that problem points to an issue with this incantation:

$node_standby->wait_for_catchup($node_cascade, 'replay',
$node_standby->lsn('replay'));

But I'm not sure, and I'm out of time to investigate for today.

--
Robert Haas
EDB: http://www.enterprisedb.com