> I don't think changing work_mem down is actually going to reduce the
> memory allocated without changing the plan to something less optimal.
> In the end, all of this is putting off the inevitable, if you get enough
> PGs going and enough requests and whatnot, you're going to start running
> out of memory again. Same if you get larger data sets that take up more
> hash table space or similar. Eventually you might need a bigger box,
> but let's try to get everything in the current box to at least be used
> first..
Yes... and indeed changing vm.overcommit_ratio to 80 does allow that
previously-failing query to execute successfully. Do you think this is
also what caused the out-of-memory error we saw today just when a
transaction was initiated?
Regards,
Matt