On 2016-08-07 14:46:06 -0400, Tom Lane wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
> > I think the whole idea of a fast temporary table is that there are no
> > catalog entries. If there are no catalog entries, then dependencies
> > are not visible. If there ARE catalog entries, to what do they refer?
> > Without a pg_class entry for the table, there's no table OID upon
> > which to depend.
>
> TBH, I think that the chances of such a design getting committed are
> not distinguishable from zero. Tables have to have OIDs; there is just
> too much code that assumes that. And I seriously doubt that it will
> work (for any large value of "work") without catalog entries.
That seems a bit too defeatist. It's obviously not a small change to get
there - and I don't think the patch upthread is really attacking the
relevant problems yet - but saying that we'll never have temp tables
without pg_class/pg_depend bloat seems to be pretty close to just giving
up. Having 8 byte oids (as explicit columns instead of magic? Or just
oid64?) and then reserving ranges for temp objects stored in a local
memory seems to be feasible. The pinning problem could potentially be
solved by "session lifetime" pins in pg_depend, which prevents dependent
objects being dropped. Obviously that's just spitballing; but I think
the problem is too big to just give up.
Andres
