Robert Haas <robertmhaas@gmail.com> writes:
> I spent a significant chunk of today burning through roughly 2^31 XIDs
> just to see what would happen. ...
> 2. Once you get to the point where you start to emit errors when
> attempting to assign an XID, you can still run plain old VACUUM
> because it doesn't consume an XID ... except that if it tries to
> truncate the relation, then it will take AccessExclusiveLock, which
> has to be logged, which forces an XID assignment, which makes VACUUM
> fail.
Yeah. I tripped over that earlier this week in connection with the
REINDEX business: taking an AEL only forces XID assignment when
wal_level is above "minimal", so it's easy to come to the wrong
conclusions depending on your test environment. I suspect that
previous testing of wraparound behavior (yes there has been some)
didn't see this effect because the default wal_level didn't use to
cause it to happen. But anyway, it's there now and I agree we'd
better do something about it.
My brain is too fried from release-note-writing to have any trustworthy
opinion right now about whether your patch is the best way.
regards, tom lane
