Date: Fri, 30 Mar 2007 04:46:11 -0600
From: Michael Fuhr <mike@fuhr.org>
To: Gaetano Mendola <mendola@bigfoot.com>
Cc: "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>,
	Claus Guttesen <kometen@gmail.com>
Subject: Re: Wrong plan sequential scan instead of an index one
Message-ID: <20070330104611.GA1540@winnie.fuhr.org>
References: <460CD941.6080303@bigfoot.com>
	<b41c75520703300251oe916648r75c16c1bf70cd10@mail.gmail.com>
	<460CE19A.3060909@bigfoot.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <460CE19A.3060909@bigfoot.com>

On Fri, Mar 30, 2007 at 12:08:26PM +0200, Gaetano Mendola wrote:
> Claus Guttesen wrote:
> > Try changing random_page_cost from the default 4 to 2 in postgresql.conf:
> > 
> > random_page_cost = 2
> 
> I have tuned that number already at 2.5, lowering it to 2 doesn't change
> the plan.

The following 19-fold overestimate is influencing the rest of the
plan:

  ->  Seq Scan on l_pvcp  (cost=0.00..2.17 rows=19 width=4) (actual time=0.066..0.081 rows=1 loops=1)
        Filter: (value ~~* '%pi%'::text)

Have you tried increasing the statistics target on l_pvcp.value?
I ran your queries against canned data in 8.2.3 and better statistics
resulted in more accurate row count estimates for this and other
parts of the plan.  I don't recall if estimates for non-leading-character
matches in earlier versions can benefit from better statistics.

-- 
Michael Fuhr