On Mon, Mar 2, 2015 at 2:19 PM, David Kubečka <kubecka.dav@gmail.com> wrote:
> The question is why optimizer, or rather the cost estimator, produced so
> much different estimates upon very small change in input. Moreover it seems
> that the main culprit of bad estimates isn't actually directly related to
> outer table, but rather to inner table. Just compare estimates for the two
> index scans:
>
> With 'random_fk_dupl':
> -> Index Scan using facts_fk_idx on facts (cost=0.42..5.75
> rows=100 width=15) (actual time=0.009..0.117 rows=98 loops=100)
> With 'random_fk_uniq':
> -> Index Scan using facts_fk_idx on facts (cost=0.42..214.26
> rows=100 width=15) (actual time=0.007..0.109 rows=98 loops=100)
Whoa. That's pretty dramatic. Am I correctly understanding that the
two tables contain *exactly* the same data? Could the issue be that
the optimizer thinks that one of the tables is thought to have a
higher logical-to-physical correlation than the other (see
pg_stats.correlation)?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company