Martijn van Oosterhout <kleptog@svana.org> writes:
> Disclaimer: The Linux OOM killer has never killed the wrong process for
> me, so I don't have any bad experiences with overcommit.
You haven't tried real hard. What I've seen recently when I do something
that makes a PG backend go overboard is that the kernel zaps both the
misbehaving backend and the bgwriter process. No idea what it's got
against the bgwriter, but the behavior's been pretty consistent under
recent Fedora 4 kernels ...
(This is on a development machine, not a server, so I'm not particularly
worried by leaving the default overcommit policy in place. I wouldn't
do that on a server --- but it's not worth my time to change it on a
devel machine.)
regards, tom lane