2011/2/26 PostgreSQL - Hans-Jürgen Schönig <postgres@cybertec.at>:
> what we are trying to do is to explicitly store column correlations. so, a histogram for (a, b) correlation and so
on.
>
The problem is that we haven't figured out how to usefully store a
histogram for <a,b>. Consider the oft-quoted example of a
<city,postal-code> -- or <city,zip code> for Americans. A histogram
of the tuple is just the same as a histogram on the city. It doesn't
tell you how much extra selectivity the postal code or zip code gives
you. And if you happen to store a histogram of <postal code, city> by
mistake then it doesn't tell you anything at all.
We need a data structure that lets us answer the bayesian question
"given a city of New York how selective is zip-code = 02139". I don't
know what that data structure would be.
Heikki and I had a wacky hand-crafted 2D histogram data structure that
I suspect doesn't actually work. And someone else did some research on
list and came up with a fancy sounding name of a statistics concept
that might be what we want.
--
greg