I wrote:
> You've still got a nasty join-size estimation error:
>> -> Nested Loop (cost=6.18..1939.43 rows=411736 width=8) (actual
>> time=0.203..3.487 rows=35 loops=1)
> It's not apparent why that's so far off ...
What PG version is this, anyway? It strikes me that this estimation
error might have something with the eqjoinsel bugs that we repaired
in 9.0.5. I'm not having any luck reproducing such a bogus estimate
with current code, either, though that may just mean you've omitted
some critical info about how the tables are set up.
regards, tom lane