On Wed, 16 Oct 2019 at 05:05, Jeremy Finzel <finz...@gmail.com> wrote: > But perhaps it would be worth exploring if there could be more detailed stats > on physical vs logical correlation, such as when ANALYZE takes its samples, > noting physical locations as well as logical values, and allowing the > correlation to take account of that more detailed analysis. Of course, > sounds like a huge amount of work with uncertain benefits.
Yes. I think improving the statistics could be beneficial. It does appear like the single value for the entire column is not fine-grained enough for your use case. > As the docs state, I do believe that the only use case that will work really > well for BRIN is either a truly insert-only table which is never pruned (an > idea I dislike as a DBA who wants us to keep OLTP data trim and implement > data retention policies :), or a table which is routinely CLUSTERed! Have you considered partitioning the table so that the retention policy could be implemented by dropping a partition rather than performing a bulk DELETE? -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services