Re: Snapshot management, final

Tom Lane wrote:
> Alvaro Herrera <alvherre@commandprompt.com> writes:
> > Tom Lane wrote:
> >> Shouldn't UnregisterSnapshot insist that s_level be equal to current
> >> xact nest level?
>
> > It can't check that; consider
>
> > begin;
> > savepoint foo;
> > declare cur cursor for select (1), (2), (3);
> > savepoint bar;
> > close cur;
> > commit;
>
> Hmm ... but that "close" can't unregister the snapshot immediately,
> because you'd lose if the 2nd savepoint gets rolled back, no?  Is the
> handling of this case even correct at the moment?

No, CLOSE is not rolled back:

alvherre=# begin;
BEGIN
alvherre=# savepoint foo;
SAVEPOINT
alvherre=# declare cur cursor for select (1), (2), (3);
DECLARE CURSOR
alvherre=# savepoint bar;
SAVEPOINT
alvherre=# close cur;
CLOSE CURSOR
alvherre=# rollback to bar;
ROLLBACK
alvherre=# fetch all from cur;
ERREUR:  le curseur « cur » n'existe pas

Maybe this is possible to fix, but again I think it's outside the scope
of this patch.

> ISTM correct handling of this example would require that the "close"
> not really discard the snap until commit.  Then, given proper ordering
> of the cleanup operations at commit, you might be able to still have the
> cross-check about s_level in UnregisterSnapshot.  (IOW, maybe having
> snapshot cleanup be late in the commit sequence wasn't such a good
> choice...)

Right -- I'll move them earlier.  I don't think it's trivial to fix the
un-rollback-ability of CLOSE however.

--
Alvaro Herrera                                http://www.CommandPrompt.com/
PostgreSQL Replication, Consulting, Custom Development, 24x7 support