čt 25. 4. 2024 v 8:52 odesílatel Andrey M. Borodin <x4mmm@yandex-team.ru> napsal:
> On 25 Apr 2024, at 11:12, Pavel Stehule <pavel.stehule@gmail.com> wrote: > > yesterday, I had to fix strange issue on standby server
It’s not just broken reading, if this standby is promoted in HA cluster - this would lead to data loss. Recently I’ve observed some lost heap updates ofter OOM-ing cluster on 14.11. This might be unrelated most probably, but I’ll post a link here, just in case [0]. In February and March we had 3 clusters with similar problem, and this is unusually big number for us in just 2 months.
Can you check LSN of blocks with corrupted tuples with pageinpsect on primary and on standby? I suspect they are frozen on primary, but have usual xmin on standby.
Unfortunately, I have not direct access to backup, so I am not able to test it. But VACUUM FREEZE DISABLE_PAGE_SKIPPING on master didn't help