Date: Sun, 23 Oct 2005 12:23:33 +0200
From: "Steinar H. Gunderson" <sgunderson@bigfoot.com>
To: pgsql-performance@postgresql.org
Subject: Re: Materializing a sequential scan
Message-ID: <20051023102333.GA9356@samfundet.no>
References: <20051019174544.GA32303@samfundet.no>
	<4248.1129783045@sss.pgh.pa.us>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <4248.1129783045@sss.pgh.pa.us>
User-Agent: Mutt/1.5.9i

On Thu, Oct 20, 2005 at 12:37:25AM -0400, Tom Lane wrote:
> That mdb_gruppekobling_transitiv_tillukning function looks awfully
> grotty ... how many rows does it return, and how long does it take to
> run by itself?  How often does its temp table get vacuumed?  A quick
> band-aid might be to use TRUNCATE instead of DELETE FROM to clean the
> table ... but if I were you I'd try to rewrite the function entirely.

I've verified that it indeed does use 20ms more for every run without a
VACUUM, but it shouldn't really matter -- and I guess it will go away once
somebody teaches plpgsql about not caching OIDs for CREATE TEMPORARY TABLE.
:-)

In any case, I still can't understand why it picks the plan it does; what's
up with the materialized seqscan, and where do the four million rows come
from? 7.4 estimates ~52000 disk page fetches for the same query, so surely
there must be a better plan than four million :-)

/* Steinar */
-- 
Homepage: http://www.sesse.net/