> > set enable_hashjoin = false;
>
> BINGO!
I'm not so sure about that. Your dataset seems to have changed fairly
significantly since the last test.
> -> Index Scan using feature_pkey on feature f (cost=0.00..134592.43 rows=47912 width=39) (actual time=0.46..502.50 rows=431 loops=1)
Notice it only pulled out 431 rows where prior runs pulled out several
thousand (~13000). I think what really happened was something came
along and deleted a bunch of stuff, then vacuum ran.