On Fri, 2025-02-21 at 21:00 -0500, Tom Lane wrote:
> I think we might indeed want that, but it doesn't seem to be the
> explanation for the buildfarm failures, because the diffs look
> to be consistent across runs which you'd not expect from
> autovacuum-driven changes. I suspect that the problem is that
> pg_dump is interpreting old-version stats in some way that doesn't
> match up with what we get from restoring the dump.
I agree it doesn't explain the failures.
> I did experiment with the attached very-quick-n-dirty patch, which
> should succeed in suppressing autovacuum in both the old and new
> versions if I understand the code correctly (which I might well not).
> It made no difference at all in the dump diffs ...
In 002_pg_upgrade.pl, I disabled autovacuum and restarted after the
regression run. In other words, in the old cluster, autovacuum did have
a chance to run, just not after the first dumpall.
Regards,
Jeff Davis
