On Tue, May 07, 2019 at 06:45:36PM -0400, Tom Lane wrote:
> Yeah. CREATE INDEX CONCURRENTLY has always had a deadlock hazard,
> so it's hardly surprising that REINDEX CONCURRENTLY does too.
> I don't think that fixing that is in-scope for v12, even if we had
> an idea how to do it, which we don't.
The most straight-forward approach I can think of would be to
determine if non-transactional commands taking a lock on a table can
be safely skipped or not when checking for older snapshots than the
minimum where the index is marked as valid. That's quite complex to
target v12, so I agree to keep it out of the stability work.
> We do need to fix the wrong-lock-level problem of course, but
> that seems straightforward.
Sure.
--
Michael