Re: Adding REPACK [concurrently]

Andres Freund <andres(at)anarazel(dot)de> wrote:

> On 2026-04-15 16:50:11 +0200, Antonin Houska wrote:
> > I thought of a "hypothetical graph", which would include the to-be-granted
> > lock, but the major issue is that it will not work correctly without the
> > locking the LMGR's LW locks we do in CheckDeadLock():
> >
> > for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
> > LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
> >
> > And obviously, doing this each time we want to insert a lock into the queue
> > would be bad for performance.
>
> Hence my suggestion to do this as part of the deadlock check. Then we don't do
> this unnecessary work outside of the case where we actually need it.

> That does need to deal with the case of the deadlock check running first in
> the backend doing repack, but that's not that hard - I think it'd be good
> enough to set its deadlock timeout temporarily to a higher value. The backend
> *should* still run the deadlock detector, because it could probably still get
> into a deadlock (e.g. due to a pg_class access or something).

Yes, the question is when we should run that check. I thought that it should
happen during each lock acquisition, and that made me worried about
performance. AFAIU you suggest REPACK to do something like:

1. Acquire ShareUpdateExclusiveLock

2. Perform the "enhanced check" to see if a future request for
AccessExclusiveLock can trigger a deadlock.

3. Do major part of the work (copy the table, build indexes, ...)

4. Request AccessExclusive lock.

5. Finish the work (process the remaining concurrent changes and swap the
table files).

This makes me concerned that if another session does

BEGIN;
TABLE t;

between steps 2 and 4, and something like

-- Get AccessExclusiveLock
ALTER TABLE t ADD COLUMN j int;

just after step 4, the two session will probably up in a deadlock anyway. In
other words, even if REPACK does the check early, it does not prevent other
sessions from getting in the way.

Maybe I'm still missing something.

--
Antonin Houska
Web: https://www.cybertec-postgresql.com