asyncore.py:
-----------

    def add_channel(self, map=None):
        #self.log_info('adding channel %s' % self)
        if map is None:
            map = self._map
        map[self._fileno] = self

As we see, it add itself to a map, creating unnecessary refences to 'self'.

Such code should be rewritten via weakref. Now it's unknown when object will garbage collected. For example, if someone forget to call close(), object will stuck, eating file descriptor and memory...

When using weakref, we may guarantee (via callback fcuntion), that we call close() when last reference to object has been lost. Also, such behaviour guarantee, that object will be garbage collected when last user's reference has gone.

To approve my thoughts, see code:
------------------------------

class MyServer(asyncore.dispatcher):
    def __init__(self, listenaddr):
        asyncore.dispatcher.__init__(self)
        self.create_socket(socket.AF_INET, socket.SOCK_STREAM)
        self.set_reuse_addr()
        self.bind(listenaddr)
        self.listen(5)
    def handle_accept(self):
        while 1:
            r=self.accept()
            if r is None:
                break
            my_conn_handler(r[0])

As we see, we create a new instance via my_conn_handler(), and we do not store reference to it, but object will not be garbage collected. It's unexpected behaviour.

Such code should be rewritten via weakref.

Can you write a patch?

sorry, forgot
"import weakref"

I'd be fine with this. My only concern are performances.
I've tried this:
http://code.google.com/p/pyftpdlib/issues/attachmentText?id=152&aid=-7106494857544071944&name=bench.py&token=bd350bbd6909c7c2a70da55db15d24ed

Results:

plain dict: 722.26 Mb/sec
WeakValueDictionary: 659.76 Mb/sec

Well, isn’t WeakRefDictionary faster than a dict here?

WeakValue*

Éric, the weak dicts are implemented in pure Python while built-in dicts are in C. That can make quite a difference.

Oh, I guess I didn’t understand at all what the numbers meant. I shouldn’t have compared the values and assumed that less was better, my bad.

Unless there's a way to automatically call close() when a dispatcher instance is no longer referenced (and I can't think of anything to do that) I'd say we better close this as rejected.

Closing out as per msg128969.