On Saturday 17 July 2010 01:53:24 Robert Haas wrote:
> On Jul 16, 2010, at 6:01 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> > Andres Freund <andres@anarazel.de> writes:
> >> What could the join removal path (and similar places) *possibly* do
> >> against such a case? Without stopping to use SnapshotNow I dont see any
> >> way :-(
> >
> > But the planner, along with most of the rest of the backend, *does* use
> > SnapshotNow when examining the system catalogs.
> >
> > I share your feeling of discomfort but so far I don't see a hole in
> > Simon's argument.
Neither do I.
> > Adding a constraint should never make a
> > previously-correct plan incorrect. Removing one is a very different
> > story, but he says he's not changing that case. (Disclaimer: I have
> > not read the patch.)
>
> Perhaps we should start by deciding whether Andres' case is a bug in the
> first place, and then we can argue about whether it's a join-removal bug,
> a lock-weakening bug, or a preexisting bug.
The case where its possible to produce such a case *after* having used/locked
all participating relations is new I think.
Being able to create invalid results by doing DDL in another connection on
not-yet-used tables is at least as old as constraint exclusion. Its a bit
easier to work around, but thats it.
So I personally would not consider this patch as having a bug anymore
(thinking helps...).
Whether the general issue is a bug or a to-be-more-exhausitive-documented-
gotcha I have no idea. I know two people having hit it in production - I dont
think its a that common issue though. Starting with the fact that not that
many people use serializable.
Just to help me: The primary reasons for using SnapshotNow is speed and in
some cases correctness (referential integrity). Right? Any other reasons?
Andres
