Robert Haas <robertmhaas@gmail.com> writes:
> Hmm. ProcessInterrupts() signals some FATAL errors while the
> connection is idle, and rumor has it that that works: the client
> doesn't immediately read the FATAL error, but the next time it sends a
> query, it tries to read from the connection and sees the FATAL error
> at that time. I wonder why that's not working here.
A likely theory is that the kernel is reporting failure to libpq's
send() because the other side of the connection is already gone.
This would be timing-dependent of course.
regards, tom lane