"Matthew Terenzio" <mterenzio@gmail.com> writes:
> here is EXPLAIN ANALYZE . I really need to work on my skills at analyzing
> these:
> Limit (cost=105505.78..105505.83 rows=20 width=655) (actual
> time=74806.973..74807.037 rows=20 loops=1)
> -> Sort (cost=105505.78..105555.44 rows=19861 width=655) (actual
> time=74806.968..74806.989 rows=20 loops=1)
> Sort Key: stories."timestamp"
> -> Nested Loop (cost=0.00..90497.94 rows=19861 width=655)
> (actual time=720.251..74798.672 rows=680 loops=1)
> -> Function Scan on q (cost=0.00..12.50 rows=1000
> width=32) (actual time=0.013..0.017 rows=1 loops=1)
> -> Index Scan using description_index on stories
> (cost=0.00..90.14 rows=20 width=623) (actual time=700.633..63243.713
> rows=680 loops=1)
> Index Cond: (stories.vectors @@ "outer".q)
> Total runtime: 74847.177 ms
Huh. The plan looks fine --- I had thought maybe the optimizer was
dropping the ball, but this seems to be more or less what you need.
The indexscan seems awfully slow though.
The only thought I have to offer is that you're apparently using quite
an old version of Postgres --- the 1000-row estimate for a scalar
function scan would only have happened in 8.0 or before. Perhaps
updating to something newer would help. I'm not sure if there are any
big performance improvements in GIST indexes per se, but in 8.2 or
8.3 you'd have the option to switch to a GIN index instead. If this
table is read-mostly then that'd be a win.
regards, tom lane