Given that some time has passed I'd like to raise this again. Any chance
of this making it back to 16 to fix the regression?
We'll be using 16 for the better part of a year still and we're having
to resort to some annoying workarounds for this.
Regards,
Bennie
On 2024/11/10 18:17, Tom Lane
Bennie Swart writes:
> -- join elimination fails
> -- expect both b and c to be eliminated, but b remains
> explain (costs off)
> select a.*
> from foo a
> left join foo b on (b.id1, b.id2) = (a.id1, a.id2)
> left join foo c on (c.id1, c.id2) = (a.id1, b.id2);
>
We are encountering this issue which results in poor planning for some
views.
Some examples to illustrate the issue:
-- setup
create table foo as
select id1, id2
from generate_series(1, 100) id1,
generate_series(1, 100) id2;
alter table foo add unique (id1, id2);
-- join elimina
The attached patch tries to fix a corner case where attr_needed of an inner
relation of an OJ contains the join relid only because another,
already-removed OJ, needed some of its attributes. The unnecessary presence of
the join relid in attr_needed can prevent the planner from further join
removals
