Re: Cross-column statistics revisited

Robert Haas wrote:
>> I think the real question is: what other kinds of correlation might
>> people be interested in representing?
> 
> Yes, or to phrase that another way: What kinds of queries are being
> poorly optimized now and why?

The one that affects our largest tables are ones where we
have an address (or other geo-data) clustered by zip, but
with other columns (city, county, state, school-zone, police
beat, etc) used in queries.

Postgres considers those unclustered (correlation 0 in the stats),
despite all rows for a given value residing on the same few pages.


I could imagine that this could be handled by either some cross-column
correlation (each zip has only 1-2 cities); or by an enhanced
single-column statistic (even though cities aren't sorted alphabetically,
all rows on a page tend to refer to the same city).