Python 3.4 asyncio gc bug

#!/usr/bin/env python

"""
This is an example of what seems to be a garbage collection bug in Python 3.4.0
that does not exist in Python 3.3.3.

The example uses asyncio to create a producer that multiplexes to N consumers.
On my Mac, with N=100, 38 consumers are incorrectly garbage collected and only execute
once, dropping the surviving consumer count to 62.

Setting DO_GC to True fixes this for some unknown reason.

This problem does NOT seem to exist in Python 3.3.3 with asyncio-0.41.

Setting PYTHONASYNCIODEBUG=1 changes the number 62 to 78.
"""

import asyncio
import gc
import weakref

N = 100

Q_SET = weakref.WeakSet()

# setting this to True fixes the problem. But it's clearly a kludge.

DO_GC = False

CONSUMER_COUNT = 0

class Q_as_callable(object):
    def __init__(self, n):
        self.q = asyncio.Queue()
        self.n = n

    def __call__(self):
        return (yield from self.q.get())

def new_get_next_msg_f(x):
    q = Q_as_callable(x)
    Q_SET.add(q)
    return q

def send_to_qs(item):
    for q in Q_SET:
        q.q.put_nowait(item)

def producer():
    global CONSUMER_COUNT
    n = 0
    while 1:
        CONSUMER_COUNT = 0
        send_to_qs(n)
        yield from asyncio.sleep(1)
        print("surviving consumers: %d" % CONSUMER_COUNT)
        n += 1

def consumer(x):
    global CONSUMER_COUNT
    next_msg = new_get_next_msg_f(x)
    while 1:
        CONSUMER_COUNT += 1
        v = yield from next_msg()

def main():
    asyncio.async(producer())
    for i in range(N):
        asyncio.async(consumer(i))
    ## if you uncomment out the gc.collect, no consumers are lost.
    ## This does not make sense to me.
    if DO_GC:
        gc.collect()
    asyncio.get_event_loop().run_forever()

if __name__ == '__main__':
    main()