There is actually another separate issue here apart from that the planner obviously choosing the wrong plan as originally described in my last message, a plan it knows to be more expensive based on cost estimates.
Take a look at the way the filter condition is treated differently when estimating the number of returned rows when applied in different nodes.
Queries A/B:
-> Finalize HashAggregate (cost=32879.78..33102.62 rows=22285 width=12) (actual time=450.724..458.667 rows=15521 loops=1)
Group Key: b.item
Filter: (sum(b.amount) >= '1'::double precision)
Rows Removed by Filter: 48277
-> Gather ...
Query C:
-> Subquery Scan on c (cost=32768.35..33269.76 rows=7428 width=12) (actual time=456.591..475.204 rows=15521 loops=1)
Filter: (c.stock >= '1'::double precision)
Rows Removed by Filter: 48277
-> Finalize HashAggregate (cost=32768.35..32991.20 rows=22285 width=12) (actual time=456.582..468.124 rows=63798 loops=1)
Group Key: b.item
-> Gather ...
Interestingly enough the subquery scan with query C correctly accounts for the filter when estimating rows=7428, while A/B doesn't seem to account for the filter in the HasAggregate node (estimated rows=22285). This looks like a bug.
--
Sent via pgsql-performance mailing list (pgsql-performance@xxxxxxxxxxxxxx)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance
