Hi,
In two recent investigations in occasional test failures
(019_replslot_limit.pl failures, AIO rebase) the problems are somehow tied to
checkpointer.
I don't yet know if actually causally related to precisely those failures, but
when running e.g. 027_stream_regress.pl, I see phases in which many backends
are looping in RegisterSyncRequest() repeatedly, each time sleeping with
pg_usleep(10000L).
Without adding instrumentation this is completely invisible at any log
level. There's no log messages, there's no wait events, nothing.
ISTM, we should not have any loops around pg_usleep(). And shorter term, we
shouldn't have any loops around pg_usleep() that don't emit log messages / set
wait events. Therefore I propose that we "prohibit" such loops without at
least a DEBUG2 elog() or so. It's just too hard to debug.
The reason for the sync queue filling up in 027_stream_regress.pl is actually
fairly simple:
1) The test runs with shared_buffers = 1MB, leading to a small sync queue of
128 entries.
2) CheckpointWriteDelay() does pg_usleep(100000L)
ForwardSyncRequest() wakes up the checkpointer using SetLatch() if the sync
queue is more than half full.
But at least on linux and freebsd that doesn't actually interrupt pg_usleep()
anymore (due to using signalfd / kqueue rather than a signal handler). And on
all platforms the signal might arrive just before the pg_usleep() rather than
during, also not causing usleep to be interrupted.
If I shorten the sleep in CheckpointWriteDelay() the problem goes away. This
actually reduces the time for a single run of 027_stream_regress.pl on my
workstation noticably. With default sleep time it's ~32s, with shortened time
it's ~27s.
I suspect we need to do something about this concrete problem for 14 and
master, because it's certainly worse than before on linux / freebsd.
I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd
require adding a new enum value to WaitEventTimeout in 14. Which probably is
fine?
Greetings,
Andres Freund
