Re: Recent pg_rewind test failures in buildfarm

On Mon, Apr 14, 2025 at 10:58:28PM -0400, Tom Lane wrote:
> That assert appears to be several years old, and the
> 008_min_recovery_point.pl test script that's triggering it hasn't
> changed very recently either, so I'm baffled where to start digging.
> It has the odor of a timing problem, so maybe we just started hitting
> this by chance. Still ... anybody have an idea?

Likely 039549d70f6a. The proc_exit(0) calls done at this stage are
reporting that we have non-flushed stats, but we should have some
remaining around. This sequence is part of the pgstat_shutdown_hook
callback, which has been initialized by pgstat_initialize() done for
the WAL senders in BaseInit(). It looks like we are going to need to
be more aggressive with the stats flushes, so as we have nothing left
when reaching this stage of the shutdown in ProcessRepliesIfAny().

(Sorry for the fuzzy reply, I cannot dive into all the details now,
that's just a quick take on the matter so I may be wrong with my
assumptions.)
--
Michael