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

After some research, I think it actually has to do with the value of the 
'backlog' parameter when creating a Listener instance: if there are too 
many queued requests to a socket then further connection requests are 
refused;  the 'backlog' parameter appears to set the maximum size of the 
queue.

...or something like that.  I'm a novice here... :-)

Anyway, in the __init__ method of the Server class, in managers.py, the 
Server creates a Listener instance with the line

        self.listener = Listener(address=address, backlog=5)

When I change backlog to 50 instead of 5, I don't see the connection 
refused error any more.

I'm not suggesting that the backlog value should be changed---it 
probably shouldn't.  But this does at least explain why connections can 
be refused, and why they should be retried in that case.

I do still think that the patch, or something like it, should be 
applied.