Gregory Stark <stark@enterprisedb.com> writes:
> The screw case I've seen is when you have a large partitioned table where
> constraint_exclusion fails to exclude the irrelevant partitions. You're going
> to get 0 rows from all but the one partition which contains the 1 row you're
> looking for. But since each partition is clamped to 1 you end up with an
> estimate of a few hundred rows coming out of the Append node.
> The natural way to kill this is to allow fractional rows for these scans.
No, the right fix is to fix the constraint-exclusion failure.
> Alternatively we could make Append more clever about estimating the number of
> rows it produces. Somehow it could be informed of some holistic view of the
> quals on its children and how they're inter-dependent. If it's told that only
> one of its children will produce rows then it can use max() instead of sum()
> to calculate its rows estimate.
This gets back to the discussions at PGCon about needing to have a more
explicit representation of partitioning. Right now, for a
many-partition table we spend huge amounts of time deriving the expected
behavior from first principles, each time we make a plan. And even then
we can only get it right for limited cases (eg, no parameterized
queries). If the planner actually understood that a set of tables
formed a partitioned set then it'd be a lot easier and faster to get the
desired behavior, not only with respect to the rowcount estimates but
the plan's structure.
regards, tom lane
