Heikki Linnakangas <heikki.linnakangas@enterprisedb.com> writes:
> Greg Stark wrote:
>> I wonder if we should switch to keeping reltuplesperpage instead. Then a
>> partial vacuum could update it by taking the average number of tuples
>> per page forbthe pages it saw. Perhaps adjusting it to the weights
>> average between the old value and the new value based on how many pages
>> were seen.
> The pages scanned by a partial vacuum isn't a random sample of pages in
> the table. That would bias the reltuplesperpage value towards those
> pages that are updated more.
Yeah ... and it's highly likely that repeatedly-updated pages would have
more dead space than never-updated ones, so there'd be a systematic
creep towards underestimation of the total tuple count.
I think your previous sketch is right: suppress update of reltuples (and
relpages) from a partial vacuum scan, and ensure that the analyze phase
is allowed to do it instead if it happens during VACUUM ANALYZE.
regards, tom lane
