On Thu, Apr 6, 2023 at 9:11 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Richard Guo <guofenglinux@gmail.com> writes:
> > Thanks for reminding. Attached is the rebased patch, with no other
> > changes. I think the patch is ready for commit.
>
> Pushed after a little further fooling with the comments. I also had
> to rebase it over 11c2d6fdf (Parallel Hash Full Join). I think I did
> that correctly, but it's not clear to me whether any of the existing
> test cases are now doing parallelized hashed right antijoins. Might
> be worth a little more testing.
I don't see any (at least that are EXPLAINed). Wondering if we should
add some of these into join_hash.sql, but probably not before I figure
out how to make that whole file run faster...
> I think that Alvaro's concern about incorrect cost estimates may be
> misplaced. I couldn't find any obvious errors in the costing logic for
> this, given that we concluded that the early-exit runtime logic cannot
> apply. Also, when I try simply executing Richard's original test query
> (in a non-JIT build), the runtimes I get line up quite well ... maybe
> too well? ... with the cost estimates:
>
> v15 HEAD w/patch Ratio
>
> Cost estimate 173691.19 90875.33 0.52
> Actual (best of 3) 514.200 ms 268.978 ms 0.52
>
> I think the smaller differentials you guys were seeing were all about
> EXPLAIN ANALYZE overhead.
I tried the original example from the top of this thread and saw a
decent speedup from parallelism, but only if I set
min_parallel_table_scan_size=0, and otherwise it doesn't choose
Parallel Hash Right Anti Join. Same if I embiggen bar significantly.
Haven't looked yet but I wonder if there is some left/right confusion
on parallel degree computation or something like that...