On 10/29/2015 11:24 AM, Tom Lane wrote:
> Matthew Bellew <matthewb@labkey.com> writes:
>> I made have several users encounter performance problems, which all
>> seem to come down to this problem: multiplying selectivity estimates can
>> cause tuple estimates to grow very small very quickly, once the estimator
>> gets to 1 row, the planner may choose plans that are very good ONLY WHEN
>> there is exactly 1 row (maybe even O(N^large)). Unfortunately, these may
>> be the worst plans if the estimate is even slightly off (even just
>> returning 2 or 3 rows versus 1).
>
> Yeah, this is a well-known problem. There has been prior discussion along
> the same lines as you mention (only believe 1-row estimates when it's
> provably true that there's at most one row), but it hasn't looked like an
> easy change. See the pgsql-hackers archives for previous threads.
Also see Tomas's correlated stats patch submitted for 9.6.
--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com