I wrote:
> ... but I suddenly fear that we've missed a fundamental point about
> pg_clog truncation. And WAL wraparound for that matter. To wit, a
> sufficiently long-lived temp table could contain old XIDs, and there's
> no way for anyone except the owning backend to clean them out, or even
> guarantee that they're marked committed.
After further thought I believe this is OK as of 8.2, because a temp
table's relfrozenxid is tracked independently of any other's. (This
problem puts a stake through the heart of the recently-discussed idea
that a temp table might be able to get along without a globally visible
pg_class entry, however.)
But it seems that we need a band-aid for 8.1 and earlier. The simplest
fix I can think of is for vacuum not to attempt to advance the
datvacuumxid/datfrozenxid fields if it skipped over any temp tables of
other backends. That's a bit nasty, since in a database making heavy
use of temp tables, you might do a whole lot of vacuums without ever
meeting that condition. Anyone have a better idea?
regards, tom lane