On Sat, 3 Dec 2022 at 20:36, Ankit Kumar Pandey <itsankitkp@gmail.com> wrote:
> Shouldn't this be an acceptable tradeoff if someone wants to perform
> extra operation in plain old aggregates? Although I am not sure how much
> this extra memory and compute usage is considered as acceptable.
We do our best to ensure that a given executor node never uses more
than work_mem. Certainly, we still do have nodes that can exceed this
by a long way. It would be unlikely that we'd accept anything new
that could do this. Since nodeWindowAgg.c already can use up to
work_mem for the tuplestore, it does not seem unreasonable that if
there is a DISTINCT aggregate that you could use 50% of work_mem for
each, that is, providing you can code it in such a way that you only
allocate one of these at once, i.e not allocate one per DISTINCT
aggregate all at once.
David