<ogjunk-pgjedan(at)yahoo(dot)com> writes:
> Aha! set hashjoin=off did the trick.
> -> Index Scan using ix_user_url_tag_user_url_id on user_url_tag userurltag0_ (cost=0.00..157.34 rows=103 width=14) (actual time=1.223..1.281 rows=5 loops=1666)
> Index Cond: (userurltag0_.user_url_id = "outer".id)
This seems to be the problem right here: the estimate of matching rows
is off by a factor of 20, and that inflates the overall cost estimate
for this plan about the same, causing the planner to think the other way
is cheaper.
What does the pg_stats row for user_url_tag.user_url_id contain?
Have you analyzed that table recently?
regards, tom lane