I ran into an odd problem while testing the next release of PyMongo, the Python driver for MongoDB which I help develop. We’re improving its support for Gevent, so we’re of course doing additional tests that begin with:

from gevent import monkey; monkey.patch_socket()

Now, some tests rely on this patching, and some rely on not being patched. Gevent doesn’t provide an unpatch_socket, so I had a clever idea: I’ll fork a subprocess with multiprocessing, do the test there, and return its result to the parent process in a multiprocessing.Value. Then subsequent tests won’t be affected by the patching.

SUCCESS = 1
FAILURE = 0

def my_test(outcome):
    from gevent import monkey; monkey.patch_socket()
    # do the test ....
    outcome.value = SUCCESS

class Test(unittest.TestCase):
    def test(self):
        outcome = multiprocessing.Value('i', FAILURE)
        multiprocessing.Process(target=my_test, args=(outcome,)).start().join()
        self.assertEqual(SUCCESS, outcome.value)

Nice and straightforward, right? In sane operating systems this worked great. On Windows it broke horribly. When I did python setup.py test, instead of executing my_test(), multiprocessing on Windows restarted the whole test suite, which started another whole test suite, … Apparently, since Windows can’t fork(), multiprocessing re-imports your script and attempts to execute the proper function within it. If the test suite is begun with python setup.py test, then everything goes haywire. This problem with multiprocessing and unittests on Windows was discussed on the Python mailing list last February.

After some gloomy minutes, I decided to look at what patch_socket() is doing. Turns out it’s simple, so I wrote a version which allows unpatching:

def patch_socket(aggressive=True):
    """Like gevent.monkey.patch_socket(), but stores old socket attributes for
    unpatching.
    """
    from gevent import socket
    _socket = __import__('socket')

    old_attrs = {}
    for attr in (
        'socket', 'SocketType', 'create_connection', 'socketpair', 'fromfd'
    ):
        if hasattr(_socket, attr):
            old_attrs[attr] = getattr(_socket, attr)
            setattr(_socket, attr, getattr(socket, attr))

    try:
        from gevent.socket import ssl, sslerror
        old_attrs['ssl'] = _socket.ssl
        _socket.ssl = ssl
        old_attrs['sslerror'] = _socket.sslerror
        _socket.sslerror = sslerror
    except ImportError:
        if aggressive:
            try:
                del _socket.ssl
            except AttributeError:
                pass

    return old_attrs


def unpatch_socket(old_attrs):
    """Take output of patch_socket() and undo patching."""
    _socket = __import__('socket')

    for attr in old_attrs:
        if hasattr(_socket, attr):
            setattr(_socket, attr, old_attrs[attr])


def patch_dns():
    """Like gevent.monkey.patch_dns(), but stores old socket attributes for
    unpatching.
    """
    from gevent.socket import gethostbyname, getaddrinfo
    _socket = __import__('socket')

    old_attrs = {}
    old_attrs['getaddrinfo'] = _socket.getaddrinfo
    _socket.getaddrinfo = getaddrinfo
    old_attrs['gethostbyname'] = _socket.gethostbyname
    _socket.gethostbyname = gethostbyname

    return old_attrs


def unpatch_dns(old_attrs):
    """Take output of patch_dns() and undo patching."""
    _socket = __import__('socket')

    for attr in old_attrs:
        setattr(_socket, attr, old_attrs[attr])

In Gevent’s version, calling patch_socket() calls patch_dns() implicitly, in mine you must call both:

class Test(unittest.TestCase):
    def test(self):
        old_socket_attrs = patch_socket()
        old_dns_attrs = patch_dns()

        try:
            # do test ...
        finally:
            unpatch_dns(old_dns_attrs)
            unpatch_socket(old_socket_attrs)

Now I don’t need multiprocessing at all.