Re: BUG #18014: Releasing catcache entries makes schema_to_xmlschema() fail when parallel workers are used

At Sun, 16 Jul 2023 23:00:01 +0300, Alexander Lakhin <exclusion(at)gmail(dot)com> wrote in
> The parallel worker doesn't have a lock on pg_class_oid_index before
> executing the query, so it gets the lock and res == LOCKACQUIRE_OK
> (not
> LOCKACQUIRE_ALREADY_CLEAR as in a regular backend case), after that it
> processes invalidation messages (this can make the backend use a newer
> catalog snapshot), and at the end it does systable_endscan() ->
> index_close() -> UnlockRelationId() -> LockRelease()...
> Thus, on a next iteration it gets the lock anew, with the res ==
> LOCKACQUIRE_OK
> again, and all that ceremony repeated.
>
> It's not clear to me, whether this parallel worker behavior is
> expected and
> if so, what to fix to avoid the test failure.

That is, the function is not parallel-safe. In fact it is marked as
'r' in pg_proc.proparallel. So, the real question appears to be how it
ended up running in a paralell worker.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center