Palle Girgensohn <girgen(at)pingpong(dot)net> writes:
> When running
> select max("when") from login where userid='userid'
> it takes a second or two, but when adding "group by userid" the planner
> decides on using another plan, and it gets *much* faster. See example below.
It's only faster for cases where there are few or no rows for the
particular userid ...
> Number of tuples per user varies from zero to a couple of thousands.
The planner is using an intermediate estimate of 406 rows. You might be
well advised to increase the statistics target for login.userid --- with
luck that would help it to choose the right plan type for both common
and uncommon userids.
regards, tom lane