Re: benchmarking the query planner

"Robert Haas" <robertmhaas@gmail.com> writes:
> I had this idle thought too, but I didn't write it down because...

>> ought to be, but it seems like it ought to be possible to determine
>> that given a desired maximum error in the overall estimate.  I'm also
>> not very clear on what the "total frequency" computations (matchfreq2
>> and unmatchfreq2 in the current code) ought to look like if we are using
>> a variable subset of the inner list.

> ...of this exact concern, which I think is an insurmountable problem.

Maybe so.  If we stick to the other design (end both lists at a preset
frequency threshold) then the math clearly goes through the same as
before, just with num_mcvs that are determined differently.  But can
we prove anything about the maximum error added from that?
        regards, tom lane