Maxim Boguk <maxim(dot)boguk(at)gmail(dot)com> writes:
> But it seems that index scan cost for very narrow/selective conditional
> indexes is greatly overestimated at least in some cases.
I realized in connection with
http://archives.postgresql.org/pgsql-general/2012-01/msg00459.php
that btcostestimate is not correctly estimating numIndexTuples for
partial indexes. But it's impossible to tell from this amount of
information whether you're seeing an effect of that, or something else.
Can you provide a self-contained test case?
regards, tom lane