Justin Arnold <jharnold81@gmail.com> writes:
> Hey guys, having a bizarre issue I have not seen before and couldn't locate
> anything via Google (though I might not be choosing the correct phrasing).
Apparently, those last couple of rows in pg_attribute are visible to a
regular MVCC scan but not to SnapshotNow scans. This suggests
transaction wraparound or maybe broken hint-bit settings. Try doing a
VACUUM on pg_attribute, then VACUUM FREEZE pg_attribute. It's hard to
predict in advance of the experiment whether you'll end up with the rows
visible or not-visible, so you might lose the data in those columns ---
but apparently you're not using them anyway, or you'd be seeing more
stuff fall over than just pg_dump.
If you need to be sure that the rows end up visible, another line of
attack would be to do a manual no-op UPDATE on just those rows.
> We are running postgres 8.3.4 on 64 bit Red Hat kernel
As noted elsewhere, this is an old version with many known bugs.
regards, tom lane
