On Mon, Jul 25, 2022 at 11:35:12AM -0400, Tom Lane wrote:
> Noah Misch <noah@leadboat.com> writes:
> > On Mon, Jul 25, 2022 at 09:44:21AM -0400, Andrew Dunstan wrote:
> >> Perhaps we should have a guard in system_or_bail() and/or system_log()
> >> which bails if some element of @_ is undefined.
>
> +1, seeing how hard this is to diagnose.
>
> > That would be reasonable. Also reasonable to impose some long timeout, maybe
> > 10x or 100x PG_TEST_TIMEOUT_DEFAULT, on calls to those functions.
>
> Why would it need to be more than PG_TEST_TIMEOUT_DEFAULT?
We run some long commands, like the parallel_schedule runs. Those currently
use plain system(), but they probably should have used system_log() from a
logging standpoint. If they had, PG_TEST_TIMEOUT_DEFAULT would have been too
short. One could argue that anything that slow should declare its intent to
be that slow, but that argument is getting into the territory of a policy
change rather than a backstop for clearly-unintended longevity.