Jeff Janes <jeff.janes@gmail.com> writes:
> I'm trying to figure out why hash joins seem to be systematically underused
> in my hands. In the case I am immediately looking at it prefers a merge
> join with both inputs getting seq scanned and sorted, despite the hash join
> being actually 2 to 3 times faster, where inputs and intermediate working
> sets are all in memory. I normally wouldn't worry about a factor of 3
> error, but I see this a lot in many different situations. The row
> estimates are very close to actual, the errors is only in the cpu estimates.
Can you produce a test case for other people to look at?
What datatype(s) are the join keys?
> A hash join is charged cpu_tuple_cost for each inner tuple for inserting it
> into the hash table:
Doesn't seem like monkeying with that is going to account for a 3x error.
Have you tried using perf or oprofile or similar to see where the time is
actually, rather than theoretically, going?
regards, tom lane