Re: misbehaving planer?

On Fri, 2006-10-20 at 12:27 -0400, Tom Lane wrote:
> > This was a direct port from a big fat table. I agree, I'm not
> convinced that
> > the partial indexes will buy me much, but this box is so IO bound
> that the
> > planner overhead my just offset the needing to IO bigger indexes.
>
> Well, you should measure it, but I bet the planner wastes way more
> time
> considering the twenty-some indexes than is saved by avoiding one
> level
> of btree search, which is about the most you could hope for.

I note that in allpaths.c:set_plain_rel_pathlist() we consider partial
indexes before we consider constraint exclusion. We normally wouldn't
notice that but, in this case, that would be a big loss.

Is there a reason for that? check_partial_indexes() doesn't seem to have
important side-effects that are required for testing whether
relation_excluded_by_constraints()

--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com