Dimitri Fontaine escribió:
> Alvaro Herrera <alvherre@commandprompt.com> writes:
> > Dimitri Fontaine escribió:
> >> Why can't MVCC apply here? You'd have two versions of the pg_class entry
> >> that just has been CLUSTERed, and you keep the old relfilenode arround
> >> too. MVCC applies, and you teach vacuum to clean out the old file when
> >> cleaning out the no more visible tuple.
> >
> > It just doesn't work. pg_class (and various other rels) are special
> > because they are needed to bootstrap the catalog system. See
> > RelationCacheInitializePhase3. It wouldn't be possible to figure out
> > what's pg_class relfilenode until you have read it from pg_class, which
> > is precisely what we're trying to do.
>
> Well at bootstrap time I guess noone is able to disturb the system by
> placing a concurrent CLUSTER pg_class; call. Once started, do those rels
> still need to have a special behavior?
The relcache need to be bootstrapped more than once, not just at
initdb's bootstrap. (I guess you could try a breakpoint in formrdesc)
> I guess I'm being dense, will now let people in the know find a solution...
Yeah, well, Tom just posted a possible solution :-)
--
Alvaro Herrera http://www.CommandPrompt.com/
PostgreSQL Replication, Consulting, Custom Development, 24x7 support