To: Doug Eck <deck1@yahoo.com>
cc: pgsql-performance@postgresql.org
Subject: Re: Identical DB's, different execution plans 
In-reply-to: <822909.77097.qm@web45203.mail.sp1.yahoo.com> 
References: <822909.77097.qm@web45203.mail.sp1.yahoo.com>
Comments: In-reply-to Doug Eck <deck1@yahoo.com>
	message dated "Mon, 29 Sep 2008 11:18:45 -0700"
Date: Mon, 29 Sep 2008 19:20:20 -0400
Message-ID: <25755.1222730420@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Doug Eck <deck1@yahoo.com> writes:
> The new plan from the slower db:

>    ->  Index Scan using t1_uc2 on t1  (cost=0.00..25604.74 rows=204906 width=32) (actual time=0.061..327.285 rows=8438 loops=1)
>          Filter: active

This seems a bit fishy.  In the first place, with such a simple filter
condition it shouldn't be that far off on the rowcount estimate.  In
the second place, the cost estimate is more than twice what the other
server estimates to do a seqscan and sort of the same data, and the
rowcount estimate is five times as much.  So there's something really
significantly different about the t1 tables in the two cases.

The first thing you ought to do is to look at the pg_class.relpages
and reltuples entries for t1 in both databases.  What I am suspecting is
that for some reason the "slow" db has suffered a lot of bloat in that
table, leading to a corresponding increase in the cost of a seqscan.
If so, a VACUUM FULL or CLUSTER should fix it, though you'll next need
to look into why routine vacuumings weren't happening.  (It looks like
t2 may be a bit bloated as well.)

If that's not it, we'll need to probe deeper ...

			regards, tom lane