On 23/09/2025 6:11 PM, Tomas Vondra wrote:
> Hi,
>
> I kept looking at this, and unfortunately the it seems a bit worse :-(
>
> Fixing the overflow is easy enough - adding the cast does the trick.
>
> But then there's the second issue I mentioned - the loop does not adjust
> the hash_table_bytes. It updates the *space_allowed, but that's not what
> the current_space/new_space formulas use.
>
> This breaks the "balancing" as the nbatch gets decreased until
>
> nbatch <= (work_mem / BLCKSZ)
>
> while the hash table "space_allowed" is increasing. This may result in
> an *increased* memory usage :-(
>
> I also noticed the code does not clamp nbuckets properly as it should.
>
>
> The question what to do about this. If we got this a week ago, I'd just
> probably just revert a1b4f289, and then try again for PG19. After all,
> the issue it meant to address is somewhat rare.
>
> But with 18 already stamped ...
>
> I've shared these findings with the rest of the RMT, I'll see what their
> thoughts are. Of course, other opinions/suggestions are welcome.
>
>
> regards
>
Hi Tomas,
If you are going to investigate this problem more, can you also look at
the related problem:
https://www.postgresql.org/message-id/flat/52b94d5b-a135-489d-9833-2991a69ec623%40garret.ru#ebe4151f1d505bbcc32cf93b2e8a1936
I proposed the patch but got no feedback.
Best regards,
Konstantin
