On Sat, Jul 14, 2001 at 10:09:02PM -0400, Tom Lane wrote:
> Martijn van Oosterhout <kleptog@svana.org> writes:
> > Yes, you actually do the query and the show what took how long. As for
> > seperating the parent and child nodes, is that really necessary?
>
> If you're going to label the nodes individually, then yes. What I take
> it you haven't realized yet is that all the nodes get an initialize call
> at start of query execution, and they all get a shutdown call at the
> end, and the actual execution is thoroughly interleaved. You're not
> going to get any meaningful info by timing the interval from startup
> to shutdown.
No, I saw that. In fact, I've put nothing in the Init or the Shutdown stage,
because as you said, it's meaningless. Currently it starts the first time
that ExecProcNode is called and ends when a NULL is returned. The
interleaving is still a problem, but I'm working on that.
> I think you'll soon find that the really wacko estimates come from
> way-off-base row count estimates, and those in turn come from bogus
> statistics about the data. So the part of your plan that involves
> instrumenting the number of tuples flowing through each node might
> have some value. Beware of rescans however.
Well some of the estimates are way off, but the rescan are interesting
because they're being called at times when I don't expect them.
When I get something concrete I'll let you know.
--
Martijn van Oosterhout <kleptog@svana.org>
http://svana.org/kleptog/
> It would be nice if someone came up with a certification system that
> actually separated those who can barely regurgitate what they crammed over
> the last few weeks from those who command secret ninja networking powers.