Klaudie Willis <[email protected]> writes:
> I'd like to add, that when I do the same query DIRECTLY on the bigtable_y2020
> (instead of the partition parent) it does change to "index scan" again.
Yeah. I think the issue here is that add_paths_to_append_rel only
considers cheapest-total paths for the member relations. Seeing
that it's already considering a slightly ridiculous number of
parallelization options, I'm hesitant to throw in cheapest-startup
considerations as well, for fear of blowing out planning time.
Maybe the right way to improve this is to bypass add_paths_to_append_rel
entirely when there's exactly one surviving child rel, and make it
just use all the surviving paths for that child.
regards, tom lane
