Thomas Munro <thomas.munro@gmail.com> writes:
> ... It works if you're OK creating partial paths
> for everything...
Hmm. The committed patch already causes us to investigate more
paths than before, which I was okay with because it only costs
more if there's an antijoin involved --- which it seems like
there's at least a 50% chance of winning on, depending on which
table is bigger.
This:
> +#if 0
> /* If any limit was set to zero, the user doesn't want a
> parallel scan. */
> if (parallel_workers <= 0)
> return;
> +#endif
seems like it adds a lot of new paths with a lot lower chance
of win, but maybe we could tighten the conditions to improve
the odds?
regards, tom lane