Hi,
One thing that occurred to me when reading the "Ad Hoc Indexes"
thread was that PG doesn't seem to do much with tidying up common
sub-expressions (I'm not sure why I remembered about it as it's not
particularly related, strange). Anyway, as an example imagine I have a
large table that I want to do a self join on:
SELECT m1.source_ls_id, m1.movement_date, m2.movement_date FROM bcms.source_movements m1, bcms.source_movements m2
WHEREm1.source_ls_id = m2.source_ls_id AND m1.movement_date < m2.movement_date;
I get a plan that sorts the movements table twice, giving the correct
answer but taking a while to actually get it.
Merge Join (cost=58981120.56..138431232.17 rows=1498156785 width=12) Merge Cond: (m1.source_ls_id =
m2.source_ls_id) Join Filter: (m1.movement_date < m2.movement_date) -> Sort (cost=29490560.28..29889000.48
rows=159376080width=8) Sort Key: m1.source_ls_id -> Seq Scan on source_movements m1
(cost=0.00..2874586.80 rows=159376080 width=8) -> Sort (cost=29490560.28..29889000.48 rows=159376080 width=8)
Sort Key: m2.source_ls_id -> Seq Scan on source_movements m2 (cost=0.00..2874586.80
rows=159376080width=8)
The time I actually tend to notice it more is when the join is between
two expensive views; this test case is nice and easy to reason about
though.
Sam