Arthur Ramsey <arthur.ramsey@code42.com> writes:
> After a restart and seq disabled for session with no other load (same for
> plan with seq enabled).
> Bitmap Heap Scan on t_user user0_ (cost=19567.17..58623.03 rows=800678
> width=761) (actual time=0.370..0.702 rows=36 loops=1)
So the problem is that awful rowcount estimate. (I see that v12's
estimate wasn't exactly spot-on either, but it wasn't so bad as to
push the planner to use a seqscan.) Hard to form an opinion on the
cause of that when you've redacted all the details of the index
condition, unfortunately.
It could be that increasing the table's statistics target and
re-ANALYZEing would help.
regards, tom lane