Re: Issues with ON CONFLICT UPDATE and REINDEX CONCURRENTLY

Hello!

On Mon, Jan 19, 2026 at 11:59 PM Mihail Nikalayeu <mihailnikalayeu(at)gmail(dot)com>
wrote:
>It fails with the state then we have 2 indexes
(leaf_part_rri->ri_NumIndices=2) and EACH OF them have an ancestor. So,
both of them are added as arbiters.

I found the cause of the issue.

As Álvaro mentioned early - get_partition_ancestors does not uses any
caches, so, each scan invalidates catalog snapshot [0].

In our case for first index - snapshot sees data
before index_concurrently_swap committed, but for second one - after. It is
rear case, but possible.

It is not a new problem (caused by that patch set), the same may happen
before, but in little bit different way:

1) some index selected (old one) for plan
2) index_concurrently_swap committed
3) get_partition_ancestors uses new catalog snapshot and sees 0 ancestors
for the index - the same error

So, test found the issue - great. I'll prepare the fix and additonal test
later this week.

Mikhail.

[0]:
https://github.com/postgres/postgres/blob/451c43974f8e199097d97624a4952ad0973cea61/src/backend/utils/time/snapmgr.c#L409-L419