On Fri, Aug 1, 2025 at 9:10 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> My thought here is that there is not a lot you can do in v17: it's
> been out for a year and we dare not destabilize it. Take the easy
> near-guaranteed win and be happy. Additional improvements can be
> pursued at our leisure in the master branch.
I agree. I always thought that it was unlikely that anybody would
seriously argue in favor of the more ambitious approach I laid out. I
nevertheless felt obligated to present it as an option.
I find it natural to approach fixing this regression (and any other
regression with a fairly crisp problem statement) by starting from the
premise that the query should be at least as fast as it was on a
previous release, and working backwards from there. Actually meeting
that standard might not be practically achievable, for whatever
reason. But even then there is still value in understanding and
clearly communicating why it is that we fell short of that standard.
--
Peter Geoghegan
