On 14 January 2013 07:35, Stefan Andreatta <s.andreatta@synedra.com> wrote:
> The source of these troubles is the sampling method employed in
> src/backend/commands/analyze.c. Judging from Tom Lane's comment for the
> original implementation in 2004 this has never been thought to be perfect.
> Does anybody see a chance to improve that part? Should this discussion be
> taken elsewhere? Is there any input from my side that could help?
Numerous alternative algorithms exist, as this has been an area of
great interest for researchers for some time. Some alternatives may
even be objectively better than Haas & Stokes. A quick peruse through
the archives shows that Simon Riggs once attempted to introduce an
algorithm described in the paper "A Block Sampling Approach to
Distinct Value Estimation":
http://www.stat.washington.edu/research/reports/1999/tr355.pdf
However, the word on the street is that it may be worth pursuing some
of the ideas described by the literature in just the last few years.
I've often thought that this would be an interesting problem to work
on. I haven't had time to pursue it, though. You may wish to propose a
patch on the pgsql-hackers mailing list.
--
Peter Geoghegan http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services