Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> What WOULD be useful is to find a way to provide a way to configure
> work_mem per backend rather than per executor node. But that's a much
> harder problem.
I think it's mostly a planner problem: how do you deal with the fact
that that would make cost estimates for different sub-problems
interrelated? The cost of a hash, for instance, depends a lot on how
much memory you assume it can use.
regards, tom lane