I said:
> While it clearly takes some amount of time to clean out our entries in
> the PROC array, etc, this trace suggests that the cost is a lot higher
> than one would've expected. Anyone have an idea what's going on here?
Oh, never mind, I figured it out: the backends are all nice'd because
they're background processes. As long as the client side wants cycles
(which it will, until it issues a query to the next backend) it's
hard for the exiting backend(s) to get any cycles at all.
regards, tom lane