On Fri, Sep 2, 2011 at 12:45 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> column values). But GROUP BY or DISTINCT would entirely invalidate the
> column frequency statistics, which makes me think that ignoring the
> pg_statistic entry might be the thing to do. Comments?
There's a possible problem there in that you may have trouble getting
a good join selectivity estimate in cases like:
SELECT ... FROM foo LEFT JOIN (SELECT x, SUM(1) FROM bar GROUP BY 1)
ON foo.x = bar.x
My guess is that in practice, the number of rows in foo that find a
join partner here is going to be much higher than what a stats-less
join selectivity estimation is likely to come up with. You typically
don't write a query like this in the first place if you don't expect
to find matches, although I'm sure it's been done. In some cases you
might even have a foreign key relationship to work with.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company