On Aug 2, 2008, at 9:44 AM, Mark Dickinson <report@bugs.python.org>  
wrote:

>
> Mark Dickinson <dickinsm@gmail.com> added the comment:
>
>> Are you looking at the conn refused or the incref error?
>
> The connection refused error.
>
> The attached patch fixes the problem, for me.  On my machine, the
> connection refused error code was 61 rather than 10061.  With this  
> patch,
> I'm no longer seeing any hangs in test_multiprocessing.py (at least,  
> not in
> the last 500 runs :-)).  (Though I am still seeing the incref error
> occasionally.)
>
> If anyone's prepared to answer a stupid question:  I'm curious why  
> failed
> socket connections occur at all.  Is connecting to a socket generally
> considered an unreliable operation, or is there some aspect of the
> multiprocessing module that makes it potentially unreliable?
>

I believe the conn refused error is another race, the child processes  
are connecting to a manager which is shutting down/gone

Thanks again mark- when I get a chance today I'll review/test/apply  
the patch