Gregory Stark <stark@enterprisedb.com> writes:
> "Tom Lane" <tgl@sss.pgh.pa.us> writes:
>> It's been clear for quite awhile that a stats target of 10 is often
>> too low, but no one has done the legwork to establish what a more
>> reasonable tradeoff point would be.
> Any ideas on what measurements would be interesting for this?
Time to run ANALYZE on large tables, extra planning time for queries of
various complexities; versus whether you actually get a better plan or
not.
In a quick look at selfuncs.c, it appears that the worst planner hit
would be for eqjoin selectivities between non-unique columns (ie, those
having MCV lists). The number of operations is O(N^2) if there are
N values in each MCV list.
regards, tom lane