Date: Thu, 1 Jun 2006 16:33:09 -0500
From: Kenneth Marshall <ktm@it.is.rice.edu>
To: Josh Berkus <josh@agliodbs.com>
Cc: Greg Stark <gsstark@mit.edu>, Tom Lane <tgl@sss.pgh.pa.us>,
	pgsql-hackers@postgresql.org
Subject: Re: More thoughts about planner's cost estimates
Message-ID: <20060601213309.GL15172@it.is.rice.edu>
References: <14821.1149104747@sss.pgh.pa.us>
	<200606011132.22066.josh@agliodbs.com>
	<87u074g03a.fsf@stark.xeocode.com>
	<200606011350.26364.josh@agliodbs.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <200606011350.26364.josh@agliodbs.com>
User-Agent: Mutt/1.4.2i

Josh, Greg, and Tom,

I do not know how sensitive the plans will be to the correlation,
but one thought might be to map the histogram X histogram correlation
to a square grid of values. Then you can map them to an integer which
would give you 8 x 8 with binary values, a 5 x 5 with 4 values per
point, or a 4 x 4 with 8 values per point. If close is good enough,
that would be a compact way to store some histogram cross correlation
information.

Ken

On Thu, Jun 01, 2006 at 01:50:26PM -0700, Josh Berkus wrote:
> Greg, Tom,
> 
> ...
> > 2) It isn't even clear what data you're exactly looking for. Certainly
> >    "correlation" is just shorthand here and isn't what you're actually
> > looking for. 
> 
> Actually, I'd think that a correlation number estimate (0 = complete 
> uncorrelated, 1 = completely correlated) would be sufficient to improve 
> row count estimation significantly, without incurring the vast overhead of 
> histogramXhistorgram manipulation.
> ...