Alvaro Herrera <alvherre@commandprompt.com> writes:
> Excerpts from Robert Haas's message of lun mar 14 11:18:24 -0300 2011:
>> Does it really matter? What Tom was describing sounded embarassingly cheap.
That was my thought exactly. If you could even measure the added cost
of doing that, I'd be astonished. It'd be adding one comparison-and-
possible-assignment to a loop that also has to invoke a binary search
of a TID array --- a very large array, in the cases we're worried about.
I'd put the actual update of pg_statistic somewhere where it only
happened once, but I don't especially care if the stat gets computed on
each index scan.
> As Heikki says, maybe this wouldn't be an issue at all if we can do it
> during ANALYZE instead, but I don't know if that works.
I'm not convinced you can get a sufficiently good estimate from a small
subset of pages.
I actually started with the idea of having ANALYZE try to calculate
correlation for multi-column indexes the same way it now calculates it
for individual data columns, but when this idea occurred to me it just
seemed a whole lot better. Note that we could remove the correlation
calculations from ANALYZE altogether.
regards, tom lane