Alexander Lakhin <exclusion@gmail.com> writes:
> 29.03.2024 16:51, Tom Lane wrote:
>> Ouch. So what's triggering that? The intention of test_setup
>> surely is to provide a uniform starting point.
> Thanks for your attention to the issue!
> Please try the attached...
I experimented with the attached modified version of the patch,
which probes just after the relevant VACUUMs and reduces the
crankiness of ConditionalLockBufferForCleanup a bit to more nearly
approximate what we're likely to see in the buildfarm. There are
two clearly-visible effects:
1. The initial VACUUM fails to count some pages as all-visible,
presumably exactly the same ones that ConditionalLockBufferForCleanup
fails on. This feels like a bug. We still scan the pages (else
reltuples would be wrong); why would we not recognize that they are
all-visible?
2. The re-VACUUM in sanity_check.sql corrects most of the
relallvisible discrepancy, presumably because it's preferentially
going after the pages that didn't get marked the first time.
However, it's distorting reltuples. Interestingly, the distortion
is worse with less-cranky ConditionalLockBufferForCleanup.
I believe the cause of the reltuples distortion is that the
re-VACUUM will nearly always scan the last page of the relation,
which would usually contain fewer tuples than the rest, but
then it counts that page equally with the rest to compute the
new tuple density. The fewer other pages are included in that
computation, the worse the new density estimate is, accounting
for the effect that when ConditionalLockBufferForCleanup is more
prone to failure the error gets smaller.
The comments in vac_estimate_reltuples already point out that
vacuum tends to always hit the last page and claim that we
"handle that here", but it's doing nothing about the likelihood
that the last page has fewer than the normal number of tuples.
I wonder if we could adjust the density calculation to account
for that. I don't think it'd be unreasonable to just assume
that the last page is only half full. Or we could try to get
the vacuum logic to report the last-page count separately ...
I tried the patch in v16 too and got similar results, so these
are not new problems.
regards, tom lane