On 2016-11-01 09:56:45 -0400, Tom Lane wrote:
> The real problem with Oskari's theory is that it requires not merely
> busted, but positively brain-dead error handling in the shell and/or
> sqsh, ie ignoring open() failures altogether. That seems kind of
> unlikely. Still, I suspect he might be onto something --- there must
> be some reason you can reproduce the issue in production and not in
> your test bed, and number-of-open-files is as good a theory as I've
> heard.
I've seen shell code akin to
exec >16 somefile # assume fd 16 is unused
more than one :(