On Sat, Dec 22, 2012 at 12:42:43AM +0000, Simon Riggs wrote:
> On 21 December 2012 20:10, Noah Misch <noah@leadboat.com> wrote:
> > I thought of one case where we do currently forget rd_newRelfilenodeSubid:
> >
> > BEGIN;
> > TRUNCATE t;
> > SAVEPOINT save;
> > TRUNCATE t;
> > ROLLBACK TO save;
>
> That's a weird one. Aborting a subtransacton that sets it, when it was
> already set.
>
> The loss of rd_newRelfilenodeSubid in that case is deterministic, but
> tracking the full complexity of multiple relations and multiple nested
> subxids isn't worth the trouble for such rare cases [assumption].
>
> I'd go for just setting an its_too_complex flag (with better name)
> that we can use to trigger a message in COPY to say that FREEZE option
> won't be honoured. That would then be completely consistent, rather
> than the lack of deterministic behaviour that Robert rightly objects
> to.
I wouldn't bother. The behavior here is deterministic, the cause clearly
traceable to the specific commands issued. Stable software won't suddenly
miss the optimization for no visible reason.