Re: Incorrect cost for MergeAppend

On Wed, 31 Jan 2024 at 02:23, Alexander Kuzmenkov
<akuzmenkov(at)timescale(dot)com> wrote:
>
> On Tue, Jan 30, 2024 at 1:20 PM David Rowley <dgrowleyml(at)gmail(dot)com> wrote:
> > You should likely focus on trying to find a test that does not require
> > making 2 tables with 10k rows.
>
> Is 1k smallint OK? It should fit in one page. Still reproduces the
> error, and the entire test case runs in under 10 ms.

I had a go at making it a bit smaller without going dangerously close
to where the plan might change. The following seems to work.

create table ma0(a int primary key);
create table ma1() inherits (ma0);
insert into ma0 select generate_series(1, 400);
insert into ma1 select generate_series(1, 200);
analyze ma0;
analyze ma1;

explain (costs off) select * from ma0 where a < 100 order by a;

drop table ma0 cascade;

As for backpatching this. It does risk plans changing in stable
versions of PostgreSQL, which we normally try to avoid. However, it
seems quite similar to 1e731ed12a, albeit, far more long-standing.
I'm currently thinking we should backpatch this back to 12.

Does anyone disagree?

David