[Python-Dev] test_multiprocessing:test_listener_client flakiness

[Trent Nelson]

> This is a common problem.  Binding to '127.0.0.1' will bind to *only*
> that address;

Indeed.

> binding to "" will bind to *all* addresses the machine
> is known by.

Agreed again.  I believe what we're dealing with here though is a lack of clarity regarding what role the 'address' attribute exposed by multiprocess.connection.Listener should play.  The way test_listener_client() is written, it effectively treats 'address' as an end-point that can be connected to directly (irrespective of the underlying family (i.e. AF_INET, AF_UNIX, AF_PIPE)).

I believe the problems we've run into stem from the fact that the API doesn't provide any guarantees as to what 'address' represents.  The test suite assumes it always reflects a connectable end-point, which I think is more than reasonable.  Unfortunately, nothing stops us from breaking this invariant by constructing the object as Listener(family='AF_INET', address=('0.0.0.0', 0)).

How do I connect to an AF_INET Listener (i.e. SocketListener) instance whose 'address' attribute reports '0.0.0.0' as the host?  I can't.

So, for now, I think we should enforce this invariant by raising an exception in Listener.__init__() if self._socket.getsockbyname()[0] returns '0.0.0.0'.  In effect, tightening up the API such that we can guarantee  Listener.address will always represent a connectable end-point.  We can look at how to service 'listen on all available interfaces' semantics at a later date -- that adds far less value IMO than being able to depend on the said guarantee.

Thoughts?

        Trent.