What's the timing of the errors? Is there a chance that we are sending
the kill signal before the signal handling thread has actually started
*and created the named pipe*?=20
We set up the signal handling stuff pretty early, but we do seem to let
the postmaster continue it's work before it's up...
Under heavy load, a signal will typically be dropped within the first
few minutes. However, it can sometimes take a little while before the
problem happens. Thousands of the same signal to the same process may
be properly handled before one is mishandled. This is not consistant
with a problem with initial creation of the pipe.
Going back to your tests, did it ever require more than one retry?
Yes, but rarely. In a 90 hour stress test with code that allowed up to 5
calls to CallNamedPipe, I found 760 signals that required a retry. Only
one required two retries. That is why I set the number of retries to 2.
The behavior might be different if the sleep interval between retries
was changed. I used a 20 ms sleep interval between retries in all my
tests, and in the patch I sent.