Re: Eager aggregation, take 3

On Sun, Nov 10, 2024 at 7:52 PM Richard Guo <guofenglinux(at)gmail(dot)com> wrote:
> > I have similar but weaker feelings about ordered aggregates. Consider:
> >
> > explain select t1.id, array_agg(t2.v order by t3.o) from t1, t2, t3
> > where t1.id = t2.id and t2.id = t3.id group by 1;
> >
>
> It seems to me that a partially aggregated row might need to be
> combined with other partially aggregated rows after the join, if they
> belong to the same t1.id group. IIUC, this implies that we cannot
> perform partial aggregation on ordered input before the join,
> otherwise we may get incorrect results during the final aggregation
> phase.

Hmm, I think you're right. I think that if the t1.id=t2.id join is one
to one, then it would work out fine, but that need not be the case.

> Hmm, currently we only consider grouped aggregation for eager
> aggregation. For grouped aggregation, the window function's
> arguments, as well as the PARTITION BY expressions, must appear in the
> GROUP BY clause. That is to say, the depname column in the first
> query, or the n column in the second query, will not be aggregated
> into the partial groups. Instead, they will remain as they are as
> input for the WindowAgg nodes. It seems to me that this ensures
> that we're good with window functions. But maybe I'm wrong.

Unfortunately, I don't know what you mean by grouped aggregation. I
think of grouping and aggregation as synonyms, pretty much.

--
Robert Haas
EDB: http://www.enterprisedb.com