Eventloops and PyZMQ

Tornado IOLoop

Facebook’s Tornado includes an eventloop for handing poll events on filedescriptors and native sockets. We have included a small part of Tornado (specifically its ioloop), and adapted its IOStream class into ZMQStream for handling poll events on ØMQ sockets. A ZMQStream object works much like a Socket object, but instead of calling recv() directly, you register a callback with on_recv(). Callbacks can also be registered for send events with on_send().


With PyZMQ’s ioloop, you can use zmq sockets in any tornado application. You must first install PyZMQ’s IOLoop, with the ioloop.install() function:

from zmq.eventloop import ioloop

This sets the global instance of tornado.ioloop.IOLoop with the global instance of our IOLoop class. The reason this must happen is that tornado objects avoid having to pass the active IOLoop instance around by having a staticmethod IOLoop.instance(), which always returns the active instance. If PyZMQ’s IOLoop is installed after the first call to IOLoop.instance() (called in almost every tornado object constructor), then it will raise an AssertionError, because the global IOLoop instance has already been created, and proceeding would result in not all objects being associated with the right IOLoop.

It is possible to use PyZMQ sockets with tornado without calling ioloop.install(), but it is less convenient. First, you must instruct the tornado IOLoop to use the zmq poller:

from zmq.eventloop.ioloop import ZMQIOLoop

loop = ZMQIOLoop()

Then, when you instantiate tornado and ZMQStream objects, you must pass the io_loop argument to ensure that they use this loop, instead of the global instance.

This is especially useful for writing tests, such as this:

from tornado.testing import AsyncTestCase
from zmq.eventloop.ioloop import ZMQIOLoop
from zmq.eventloop.zmqstream import ZMQStream

class TestZMQBridge(AsyncTestCase):

     # Use a ZMQ-compatible I/O loop so that we can use `ZMQStream`.
     def get_new_ioloop(self):
         return ZMQIOLoop()

You can also manually install this IOLoop as the global tornado instance, with:

from zmq.eventloop.ioloop import ZMQIOLoop
loop = ZMQIOLoop()

but it will NOT be the global pyzmq IOLoop instance, so it must still be passed to your ZMQStream constructors.


ZMQStream objects do have send() and send_multipart() methods, which behaves the same way as Socket.send(), but instead of sending right away, the IOLoop will wait until socket is able to send (for instance if HWM is met, or a REQ/REP pattern prohibits sending at a certain point). Messages sent via send will also be passed to the callback registered with on_send() after sending.


ZMQStream.on_recv() is the primary method for using a ZMQStream. It registers a callback to fire with messages as they are received, which will always be multipart, even if its length is 1. You can easily use this to build things like an echo socket:

s = ctx.socket(zmq.REP)
stream = ZMQStream(s)
def echo(msg):

on_recv can also take a copy flag, just like Socket.recv(). If copy=False, then callbacks registered with on_recv will receive tracked Frame objects instead of bytes.


A callback must be registered using either ZMQStream.on_recv() or ZMQStream.on_recv_stream() before any data will be received on the underlying socket. This allows you to temporarily pause processing on a socket by setting both callbacks to None. Processing can later be resumed by restoring either callback.


ZMQStream.on_recv_stream() is just like on_recv above, but the callback will be passed both the message and the stream, rather than just the message. This is meant to make it easier to use a single callback with multiple streams.

s1 = ctx.socket(zmq.REP)
stream1 = ZMQStream(s1)

s2 = ctx.socket(zmq.REP)
stream2 = ZMQStream(s2)

def echo(stream, msg):




Sometimes with an eventloop, there can be multiple events ready on a single iteration of the loop. The flush() method allows developers to pull messages off of the queue to enforce some priority over the event loop ordering. flush pulls any pending events off of the queue. You can specify to flush only recv events, only send events, or any events, and you can specify a limit for how many events to flush in order to prevent starvation.

PyZMQ and gevent

PyZMQ ≥ ships with a gevent compatible API as zmq.green. To use it, simply:

import zmq.green as zmq

Then write your code as normal.

Socket.send/recv and zmq.Poller are gevent-aware.

In PyZMQ ≥, green.device and green.eventloop should be gevent-friendly as well.


The green device does not release the GIL, unlike the true device in zmq.core.

zmq.green.eventloop includes minimally patched IOLoop/ZMQStream in order to use the gevent-enabled Poller, so you should be able to use the ZMQStream interface in gevent apps as well, though using two eventloops simultaneously (tornado + gevent) is not recommended.


There is a known issue in gevent ≤ 1.0 or libevent, which can cause zeromq socket events to be missed. PyZMQ works around this by adding a timeout so it will not wait forever for gevent to notice events. The only known solution for this is to use gevent ≥ 1.0, which is currently at 1.0b3, and does not exhibit this behavior.

See also

zmq.green examples on GitHub.

zmq.green is simply gevent_zeromq, merged into the pyzmq project.