On 07/07/2015 06:28 AM, Marc Mamin wrote:
> Sure, but on the other hand, they are so small and quick to build
> that they seem to be a good alternative when other index types are too costly,
> even if theses indexes can't deal well with all data ranges passed as query condition.
>
> Hence it would be fine if the planner could reject these indexes in the bad cases.
Oh, sorry! I didn't realize that the planner was using the BRIN index
even when it was useless; your email wasn't clear.
The problem here is that the usefulness of BRIN indexes as a cost
calculation should take correlation into account, heavily. Can we do
that? Is correlation even part of the index costing method now? How
accurate are our correlation estimates?
--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com