On Thu, Dec 06, 2007 at 11:13:18AM -0500, Tom Lane wrote:
>Indeed, and if you've got examples where it's that far off, you should
>report them.
Yeah, the trick is to get it to a digestable test case. The basic
scenario (there are more tables & columns in the actual case) is a set
of tables partitioned by date with a number of columns in one table
referencing rows in the others:
Table A (~5bn rows / 100's of partitions)
time Bkey1 Ckey1 Bkey2 Ckey2
Table B (~1bn rows / 100's of partitions)
Bkey Bval
Table C (~.5bn rows / 100's of partitions)
Ckey Cval
Bkey and Ckey are unique, but the planner doesn't know that.
Mike Stone
