On Sun, 5 Jul 2026 at 09:17, John Naylor <johncnaylorls@gmail.com> wrote:
>
> BEGIN;
> SELECT count(*) FROM g; -- first access here: queues g's pg_temp_class row
> -- (still unflushed inside this transaction)
> SAVEPOINT s1;
> DROP TABLE g; -- removes that pending queue entry
> ROLLBACK TO s1; -- pg_class row is restored by MVCC;
> -- the queue entry is not restored
> INSERT INTO g VALUES (1); -- the table still works normally
> COMMIT;
>
> I'm starting to think the queue mechanism needs serious attention, if
> not a complete re-think.
Thanks for looking!
Yes, you're right. In that example, the pg_temp_class tuple gets lost
entirely. It really needs the same sort of sub XID tracking as the
usage/storage code, to ensure that they never get lost. I'll take a
look.
Regards,
Dean
