Re: estimating # of distinct values

> Well, but that's not 7%, thats 7x! And the theorem says 'greater or equal'
> so this is actually the minimum - you can get a much bigger difference
> with lower probability. So you can easily get an estimate that is a few
> orders off.

FWIW, based on query performance, estimates which are up to 5X off are
tolerable, and anything within 3X is considered "accurate". Above 5X
the probability of bad query plans becomes problematically high.

Of course, if you're doing cross-column stats, the accuracy of each
individual column becomes critical since estimation error could be
combiniational in the worst case (i.e. if colA is 3X and colB is 0.3X
then colA<->colB will be 9X off).

Anyway, I look forward to your experiments with stream-based estimators.

--
-- Josh Berkus
PostgreSQL Experts Inc.
http://www.pgexperts.com