Re: Fixes inconsistent behavior in vacuum when it processes multiple relations

On Wed, Jun 25, 2025 at 01:12:02PM -0400, shihao zhong wrote:
> Just want to make sure, are we not going to include my original test
> to catch the future regression? Also, could someone please let me know
> how to check if the test is stable or not?

On stable branches, you could reuse the patch I have posted upthread,
even if it is not included in the tree. On HEAD and v17, you can run
the tests of src/test/modules/injection_points/ with
--enable-injection-points set.

Making the index cleanup stable takes a certain amount of cycles
because it requires a minimum amount of data, particularly for the
btree of a TOAST index. The truncation check is halfly stable: it is
possible to check that a truncation does not happen, but it can still
be slightly unstable because the option may not trigger all the time.
So, while the test success rate is perhaps close to 90%, with this
rate going down on slower machines, it is just cheaper and more
reliable to check directly the contents of VacuumParams. Note that
your original set of tests only covered the case of multiple
relations, and it missed coverage for the TOAST relation vacuumed
after its main relation.
--
Michael