Jesper Krogh <jesper(at)krogh(dot)cc> writes:
> On 2011-02-17 23:20, Tom Lane wrote:
>> The proposed patch seems wrong to me: if we're estimating on the basis
>> of most-common-value fractions, the null_frac is already accounted for,
>> because it's not part of the MCV selectivity fractions. IOW, aren't you
>> double-counting the null fraction?
> It might be the wrong place to fix, but here it seems like we're only
> counting MCE-freqs based on non-null elements:
> http://git.postgresql.org/gitweb?p=postgresql.git;a=blob;f=src/backend/tsearch/ts_typanalyze.c;h=2654d644579fd1959282d83919474f42540ca703;hb=HEAD#l396
Hmm, you're right, and the specification in pg_statistic.h neglects to
say that. This does need work.
regards, tom lane