Re: Proposal: "query_work_mem" GUC, to distribute working memory to the query's individual operators

On Tue, Feb 11, 2025 at 10:00 AM Jeff Davis <pgsql(at)j-davis(dot)com> wrote:
>
> On Mon, 2025-02-10 at 19:09 -0800, James Hunter wrote:
> > I think it makes sense to split the work into two parts: one part
> > that
> > improves SQL execution, and a second part that improves the
> > optimizer,
> > to reflect the improvements to execution.
>
> I like the idea to store the value of work_mem in the
> path/plan/executor nodes, and use that at execution time rather than
> the GUC directly.
>
> IIUC, that would allow an extension to do what you want, right? A
> planner hook could just walk the tree and edit those values for
> individual nodes, and the executor would enforce them.

Yes, exactly!

* The Path would store "nbytes" (= the optimizer's estimate of how
much working memory a given Path will use), to allow for future
optimizer logic to consider memory usage when choosing the best Path.

* The Plan would store a copy of "nbytes," along with "work_mem," and
the executor would enforce work_mem. A "(work_mem on)" option to the
"EXPLAIN" command would display both "nbytes" and "work_mem", per Plan
node.

* Either built-in logic or an extensibility hook would set "work_mem"
on each individual Plan node, based on whatever heuristic or rule it
chooses.

Right now, my prototype sets "work_mem" inside ExecInitNode().

Thanks,
James