Andres Freund <andres@anarazel.de> writes:
> On 2022-05-08 11:28:34 -0400, Tom Lane wrote:
>> Per lapwing's latest results [1], this wasn't enough. I'm again thinking
>> we should pull the whole test from the back branches.
> That failure is different from the earlier failures though. I don't think it's
> a timing issue in the test like the deadlock check one. I rather suspect it's
> indicative of further problems in this area.
Yeah, that was my guess too.
> Potentially the known problem
> with RecoveryConflictInterrupt() running in the signal handler? I think Thomas
> has a patch for that...
Maybe; or given that it's on v10, it could be telling us about some
yet-other problem we perhaps solved since then without realizing
it needed to be back-patched.
> One failure in ~20 runs, on one animal doesn't seem worth disabling the test
> for.
No one is going to thank us for shipping a known-unstable test case.
It does nothing to fix the problem; all it will lead to is possible
failures during package builds. I have no idea whether any packagers
use "make check-world" rather than just "make check" while building.
But if they do, even fairly low-probability failures can be problematic.
(I still carry the scars I acquired while working at Red Hat and being
responsible for packaging mysql: at least back then, their test suite
was full of cases that mostly worked fine, except when getting stressed
in Red Hat's build farm. Dealing with a test suite that fails 50% of
the time under load, while trying to push out an urgent security fix,
is NOT a pleasant situation.)
I'm happy to have this test in the stable branches once we have committed
fixes that address all known problems. Until then, it will just be
a nuisance for anyone who is not a developer working on those problems.
regards, tom lane
