To: Ed Tyrrill <tyrrill_ed@emc.com>
cc: pgsql-performance@postgresql.org
Subject: Re: 
In-reply-to: <1182816598.6477.31.camel@nickel.avamar.com> 
References: <A23190A408F7094FAF446C1538222F76041E1C28@avaexch01.avamar.com>
	<7008.1182809415@sss.pgh.pa.us>
	<1182816598.6477.31.camel@nickel.avamar.com>
Comments: In-reply-to Ed Tyrrill <tyrrill_ed@emc.com>
	message dated "Mon, 25 Jun 2007 17:09:58 -0700"
Date: Mon, 25 Jun 2007 21:07:14 -0400
Message-ID: <8930.1182820034@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Ed Tyrrill <tyrrill_ed@emc.com> writes:
> It seems to me that the first plan is the optimal one for this case, but
> when the planner has more information about the table it chooses not to
> use it.  Do you think that if work_mem were higher it might choose the
> first plan again?

It's worth fooling around with work_mem just to see what happens.  The
other thing that would be interesting is to force the other plan (set
enable_mergejoin = off) just to see what the planner is costing it at.
My suspicion is that the estimated costs are pretty close.

The ANALYZE stats affect this choice only in second-order ways AFAIR.
The planner penalizes hashes if it thinks there will be a lot of
duplicate values in the inner relation, but IIRC there is also a penalty
for inner duplicates in the mergejoin cost estimate.  So I'm a bit
surprised that there'd be a change.

Can you show us the pg_stats rows for the join columns after analyzing
at target 10 and target 100?

			regards, tom lane