Re: A test for replay of regression tests

On Sat, Jan 22, 2022 at 8:48 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> Unfortunately we don't quite seem there yet:
>
> I saw a couple failures like:
> https://api.cirrus-ci.com/v1/artifact/task/5394938773897216/regress_diffs/build/testrun/recovery/t/027_stream_regress/regression.diffs
> (from https://cirrus-ci.com/task/5394938773897216?logs=check_world#L183 )
>
> -- Should succeed
> DROP TABLESPACE regress_tblspace_renamed;
> +ERROR: tablespace "regress_tblspace_renamed" is not empty
>
>
> I assume the reason we see this semi-regularly when the regression tests run
> as part of 027_stream_regress, but not in the main regression test run, is
> similar to the reloptions problem, namely that we run with a much smaller
> shared buffers.
>
> I assume what happens is that this just makes the known problem of bgwriter or
> some other process keeping open a filehandle to an already deleted relation,
> preventing the deletion to "fully" take effect, worse.

Right, I assume this would be fixed by [1]. I need to re-convince
myself of that patch's correctness and make some changes after
Robert's feedback; I'll look into committing it next week. From a
certain point of view it's now quite good that we hit this case
occasionally in CI.

[1] https://commitfest.postgresql.org/36/2962/