David Rowley <dgrowleyml@gmail.com> writes:
> On Wed, 27 Mar 2024 at 18:28, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Let's wait a bit to see if it fails in HEAD ... but if not, would
>> it be reasonable to back-patch the additional debugging output?
> I think REL_16_STABLE has told us that it's not an auto-vacuum issue.
> I'm uncertain what a few more failures in master will tell us aside
> from if reltuples == 48 is consistent or if that value is going to
> fluctuate.
> Let's give it a week and see if it fails a few more times.
We have another failure today [1] with the same symptom:
ab_a2 | 0 | -1 | | 0 | 0
- ab_a2_b1 | 0 | -1 | | 0 | 0
+ ab_a2_b1 | 0 | 48 | | 0 | 0
ab_a2_b1_a_idx | 1 | 0 | t | |
Different table, same "48" reltuples. But I have to confess that
I'd not looked closely enough at the previous failure, because
now that I have, this is well out in WTFF territory: how can
reltuples be greater than zero when relpages is zero? This can't
be a state that autovacuum would have left behind, unless it's
really seriously broken. I think we need to be looking for
explanations like "memory stomp" or "compiler bug".
regards, tom lane
[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=parula&dt=2024-03-29%2012%3A46%3A02