Re: apply_scanjoin_target_to_paths and partitionwise join

On Thu, Jan 2, 2025 at 3:58 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> I am wondering if the problem is not that the plan is slower, it's
> that for some reason the planner took a lot longer to create it.
> It's very plausible that partitionwise planning takes longer, and
> maybe we have some corner cases where the time is O(N^2) or worse.

That doesn't seem like a totally unreasonable speculation, but it
seems a little surprising that retaining the non-partitionwise paths
would fix it. True, that might let us discard a bunch of partitionwise
paths more quickly than would otherwise be possible, but I wouldn't
expect that to have an impact as dramatic as what Jakub alleged. The
thing I thought about was whether there might be some weird effects
with lots of empty partitions; or maybe with some other property of
the path like say sort keys or parallelism. For example if we couldn't
generate a partitionwise path with sort keys as good as the
non-partitionwise path had, or if we couldn't generate a parallel
partitionwise path but we could generate a parallel non-partitionwise
path. As far as I knew neither of those things are real problems, but
if they were then I believe they could pretty easily explain a large
regression.

> However, this is pure speculation without a test case, and any
> proposed fix would be even more speculative. I concur with your
> bottom line: we should insist on a public test case before deciding
> what to do about it.

Yeah.

--
Robert Haas
EDB: http://www.enterprisedb.com