Richard Kut <rkut@intelerad.com> writes:
> -> Hash (cost=49229.38..49229.38 rows=328403 width=284) (actual
> time=3342.874..3342.874 rows=328403 loops=1)
> -> Index Scan using kdx_pid on pid (cost=0.00..49229.38 rows=328403
> width=284) (actual time=0.010..2110.172 rows=328403 loops=1)
Hm, it looks like you've still got the planner's hands tied behind its
back --- specifically, I'll bet enable_seqscan = off. There's no reason
at all to use a full-table indexscan to load a hash table. Please try
it again with default planner parameter settings.
A general comment from comparing your two cases is that random_page_cost
is probably too high for the scenario you are testing, which looks to be
a case where both tables are fully cached in memory. However, before
reducing it you should ask yourself whether that's still likely to be
true in production. It's a bad idea to optimize on the basis of test
cases that are much smaller than your production scenario will be ...
regards, tom lane
