Hi
On 2022-04-16 12:13:09 -0700, Andres Freund wrote:
> What confuses me so far is what already had generated stats before
> reaching pgstat_reset_after_failure() (so that the bug could even be hit
> in t/025_stuck_on_old_timeline.pl).
I see part of a problem - in archiver stats. Even in 14 (and presumably
before), we do work that can generate archiver stats
(e.g. ReadCheckpointRecord()) before pgstat_reset_all(). It's not the
end of the world, but doesn't seem great.
But since archiver stats are fixed-numbered stats (and thus not in the
hash table), they'd not trigger the backtrace we saw here.
One thing that's interesting is that the failing tests have:
2022-04-15 12:07:48.828 UTC [675922][walreceiver][:0] FATAL: could not link file "pg_wal/xlogtemp.675922" to
"pg_wal/00000002.history":File exists
which I haven't seen locally. Looks like we have some race between
startup process and walreceiver? That seems not great. I'm a bit
confused that walreceiver and archiving are both active at the same time
in the first place - that doesn't seem right as things are set up
currently.
Greetings,
Andres Freund