On Tue, Feb 19, 2019 at 11:59 PM Abi Noda <a@abinoda.com> wrote:
Thanks Justin.
The 4ms different in the examples isn't an accurate benchmark. I'm seeing about a ~20% difference over a larger sample size. And this is on a fork of the production database.
Please show the execution plans from that larger sample, if that is the one that is most relevant.
You can "set enable_bitmapscan = off" to get rid of the bitmap scan in order to see the estimated cost and actual performance of the next-best plan (which will probably the regular index scan).