Arjen van der Meijden <acmmailing@tweakers.net> writes:
> PS, In case any of the planner-hackers are reading, here are the plans
> of the first two queries, just to see if something can be done to
> decrease the differences between them.
Increase work_mem? It's not taking the hash because it thinks it won't
fit in memory ...
There is a bug here, I'd say: the rowcount estimate ought to be the same
either way. Dunno why it's not, but will look --- I see the same
misbehavior with a toy table here.
regards, tom lane