Date: Thu, 1 Jun 2006 18:22:12 -0500
From: "Jim C. Nasby" <jnasby@pervasive.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Josh Berkus <josh@agliodbs.com>, pgsql-hackers@postgresql.org
Subject: Re: More thoughts about planner's cost estimates
Message-ID: <20060601232212.GX53487@pervasive.com>
References: <14821.1149104747@sss.pgh.pa.us> <447F23A9.1090503@agliodbs.com>
	<14438.1149189309@sss.pgh.pa.us>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <14438.1149189309@sss.pgh.pa.us>
User-Agent: Mutt/1.5.11

On Thu, Jun 01, 2006 at 03:15:09PM -0400, Tom Lane wrote:
> These would all be nice things to know, but I'm afraid it's pie in the
> sky.  We have no reasonable way to get those numbers.  (And if we could
> get them, there would be another set of problems, namely plan instability:
> the planner's choices would become very difficult to reproduce.)

Speaking of plan instability, something that's badly needed is the
ability to steer away from query plans that *might* be the most optimal,
but also will fail horribly should the cost estimates be wrong. People
generally don't care about getting the absolutely most optimal plan;
they do care about NOT getting a plan that's horribly bad. One possible
way to do this would be to have the estimator calculate a worst-case
cost to go along with the best case cost, or maybe even 3 numbers:
ideal, what we think will happen, and worst-case. I know that index
scans already have that information, at least in the form of cost for a
plan with 0 correlation and one with perfect correlation.

Does anyone have any experience developing genetic algorithms? I think
coming up with cost estimators might be an ideal use for that
technology...
-- 
Jim C. Nasby, Sr. Engineering Consultant      jnasby@pervasive.com
Pervasive Software      http://pervasive.com    work: 512-231-6117
vcard: http://jim.nasby.net/pervasive.vcf       cell: 512-569-9461