On Tue, Feb 11, 2025 at 10:34 AM Richard Guo <guofenglinux@gmail.com> wrote:
> Yeah, I also feel that the virtual generated columns should adhere to
> outer join semantics, rather than being unconditionally replaced by
> the generation expressions. But maybe I'm wrong.
>
> If that's the case, this incorrect-result issue isn't limited to
> constant expressions; it could also occur with non-strict ones.
It seems that outer-join removal does not work well with virtual
generated columns.
create table t (a int, b int);
create table vt (a int primary key, b int generated always as (a * 2));
explain (costs off)
select t.a from t left join vt on t.a = vt.a where coalesce(vt.b, 1) = 1;
QUERY PLAN
---------------
Seq Scan on t
(1 row)
This plan does not seem correct to me. The inner-rel attribute 'vt.b'
is used above the join, which means the join should not be removed.
explain (costs off)
select t.a from t left join vt on t.a = vt.a where coalesce(vt.b, 1) =
1 or t.a is null;
server closed the connection unexpectedly
For this query, an Assert in remove_rel_from_query() is hit.
I haven't looked into the details yet, but I suspect that both of
these issues are caused by our failure to mark the correct nullingrel
bits for the virtual generated columns.
Thanks
Richard
