Re: planner with index scan cost way off actual cost,

"Jim C. Nasby" <jnasby 'at' pervasive.com> writes:

> On Tue, Mar 21, 2006 at 02:03:19PM +0100, Guillaume Cottenceau wrote:
> > "Jim C. Nasby" <jnasby 'at' pervasive.com> writes:
> >
> > > On Tue, Mar 21, 2006 at 10:40:45PM +1200, Mark Kirkwood wrote:
> > > > I was going to recommend higher - but not knowing what else was running,
> > > > kept it to quite conservative :-)... and given he's running java, the
> > > > JVM could easily eat 512M all by itself!
> > >
> > > Oh, didn't pick up on java being in the mix. Yeah, it can be a real pig.
> > > I think people often place too much emphasis on having a seperate
> > > application server, but in the case of java you often have no choice.
> >
> > Fortunately the servers use 2G or 4G of memory, only my test
> > machine had 1G, as I believe I precised in a message; so I'm
> > definitely going to use Mark's advices to enlarge a lot the
> > shared buffers. Btw, what about sort_mem? I have seen it only
> > little referenced in the documentation.
>
> The biggest issue with setting work_mem (you're not doing current
> development on 7.4 are you?) is ensuring that you don't push the server

Yes, we use 7.4.5 actually, because "it just works", so production
wants to first deal with all the things that don't work before
upgrading. I have recently discovered about the background writer
of 8.x which could be a supplementary reason to push for an
ugprade though.

> into swapping. Remember that work_mem controls how much memory can be
> used for EACH sort or hash (maybe others) operation. Each query can
> consume multiples of work_mem (since it can do multiple sorts, for
> example), and of course each backend could be running a query at the
> same time. Because of all this it's pretty difficult to make work_mem
> recomendations without knowing a lot more about your environment.

Ok, I see. Thanks for the info!

--
Guillaume Cottenceau