On Wed, 2006-06-07 at 16:56 -0400, Tom Lane wrote:
> Simon Riggs <simon@2ndquadrant.com> writes:
> > On Wed, 2006-06-07 at 15:32 -0400, Tom Lane wrote:
> >> I propose we revert this patch and think about an interrupt-driven
> >> sampling method instead.
>
> > I don't have much more faith in crazy scheme No.2 either. (Mine or
> > yours...)
>
> > Can we just have an option to avoid the timing altogether, please? I
> > don't want to have long discussions about instrumentation, I just want a
> > reasonably useful EXPLAIN ANALYZE in a reasonable amount of time - one
> > that we never, ever have to doubt whether the sampling works correctly
> > on a Miasmic-367 with HyperKooling.
>
> Frankly, I think the pre-existing version of EXPLAIN ANALYZE is fine.
I respect everybody's opinion, yours doubly so, as you know. But with
this current opinion, on this occasion, I believe you are in the
minority of those who have spent some time asking for EXPLAIN ANALYSEs
and have not received them because the run time is unacceptable.
> People have been hyperventilating about the timing overhead but I think
> that it's perfectly acceptable as-is.
The phrase hyperventilating is subjective. From my side, I could use the
phrase explaining-the-obvious and potentially cause rankle also. We
should be able to discuss things without that.
> Certainly the removal of timing
> is not going to convert an intolerable EXPLAIN ANALYZE runtime into an
> acceptable one;
I disagree, as have others.
> what it *is* likely to do is let you be misled about
> which part of the query is the problem.
A full EXPLAIN ANALYZE is always desirable - we agree on that. The
question is what we do when one is not available. Guessing leaves you
much more open to being misled. I'd like an option that we can take when
a full EXPLAIN ANALYZE is not possible, one that doesn't give different
results on different CPUs/OS.
We need this to help people who are in difficulty. This isn't a vanity
feature or a they-do-it-so-we-should-also. It's for support.
-- Simon Riggs EnterpriseDB http://www.enterprisedb.com