On 18/03/2026 12:42, Alexander Kuzmenkov wrote:
> On Wed, Mar 18, 2026 at 7:59 AM Michael Paquier <michael@paquier.xyz
> <mailto:michael@paquier.xyz>> wrote:
>
> Hmm. If I take this SQL sequence independently or with an
> installcheck, the one-page VACUUM path is taken during the final
> INSERT, but that's not the case of a `make check`. Could this be made
> more stable? I have not spent a lot of time on it, so I may be
> missing something obvious, of course.
>
>
> I think this might be caused by "make check" running many tests in
> parallel, so the deleting transaction is visible to some snapshots, and
> the cleanup is not done. Not sure what's the best way to improve this.
I think if you use "BEGIN; INSERT ...; ROLLBACK;" to generate the dead
tuples instead of DELETE, it will not be sensitive to concurrent
snapshots like that.
- Heikki
