Django push: Using Server-Sent Events and WebSocket with Django

http://curella.org/blog/2012/jul/17/django-push-using-server-sent-events-and-websocket/

The goal of this article is to explore and show how it's possible to implement Server-Sent Events and WebSocket with Django.

There are other implementations out there for frameworks that are designed specifically to work in event-based scenario (tornado, Node.js), and are probably better suited for implementing these kind of services.

The point of this article is not "you should use Django for that", but a more humble "here's how I made it work with Django".

The Scenario

Suppose you have a website where users can import their contacts from other services. The importing is handled off-band by some other means (most likely, a celery task), and you want to show your users a notification box when the job is done.

There are currently a few alternative technologies for pushing events to the browser: Server-Sent Events (SSE) and WebSocket.

SSEs are a simpler protocol and are easier to implement, but they provide communication only in one direction, from the server to the browser. WebSocket provides instead a bidirectional channel.

For a simple notification scenario like the above, SSEs provide just what we want, at the expenses of one long-running connection per user.

We will use Redis and its PubSub functionality as a broker between the celery task and Django's web process.

The final code of this article is available as a repository on GitHub.

Architecture

Celery Task -> redis -> Django -> Browser

Running gunicorn

Both technologies require the server to keep the connection open indefinitely.

If we'd run Django under mod_wsgi or the regular dev server, the request-response cycle will be blocked by those always-open requests.

The solution is to use gevent. I found that the simplest way to use it is to run Django under gunicorn.

Install gunicorn:

$ pip install gunicorn

Add gunicorn to your INSTALLED_APPS:

INSTALLED_APPS = (

    ...,

    'myapp',

    'gunicorn',

)

Then, I created a config file for gunicorn at config/gunicorn.

#!python

from os import environ

from gevent import monkey

monkey.patch_all()

bind = "0.0.0.0:8000"

workers = 1 # fine for dev, you probably want to increase this number in production

worker_class = "gunicorn.workers.ggevent.GeventWorker"

You can start the server with:

$ gunicorn_django -c config/gunicorn

For more info on Django on gunicorn see Django's docs on How to use Django with Gunicorn.

Server-Sent Events

The browser will issue a GET request to the url /sse/ (this path is completely arbitrary). The server will respond with a stream of data, without ever closing the connection.

The easiest way to implement SSEs is to use the django-sse package, available on PyPi.

$ pip install sse django-sse

If you want to publish via redis, django-sse requires you to specify how to connect:

settings.py:

REDIS_SSEQUEUE_CONNECTION_SETTINGS = {

    'location': 'localhost:6379',

    'db': 0,

}

django_sse provides a ready-to-use view that uses redis as message broker.

myapp/views.py:

from django.views.generic import TemplateView

from django_sse.redisqueue import RedisQueueView

class HomePage(TemplateView):

    template_name = 'index.html'

class SSE(RedisQueueView):

    pass

Hook the views up in your urls.py:

from django.conf.urls import patterns, include, url

from myapp import views

urlpatterns = patterns('',

    url(r'^sse/$', views.SSE.as_view(), name='sse'),  # this URL is arbitrary.

    url(r'^$', views.HomePage.as_view(), name='homepage'),

)

IE Polyfill

Not every browser supports SSEs (most notably, internet Explorer).

For unsupported browser, we can include a JavaScript polyfill in our page. There are many polyfills available out there, but I've choose to use eventsource.jsbecause it's close to the original API and it looks actively maintained.

After including the polyfill in our HTML we can set up our callback functions on DOMReady. Here I've also uses jQuery for simplicity.

<!doctype html>

<html>

<head>

  <meta charset="utf-8">

  <title>My App</title>

</head>

<body>

  <script src="//ajax.googleapis.com/ajax/libs/jquery/1.7.1/jquery.min.js"></script>

  <script>window.jQuery || document.write('<script src="{{ STATIC_URL }}js/libs/jquery-1.7.1.min.js"><\/script>')</script>

  <script src="{{ STATIC_URL }}js/libs/eventsource.js"></script>

<script>

$().ready(function() {

  var source = new EventSource('/sse/'); // of course this must match the endpoint in your urlconf

  function log() {

    console.log(arguments);

  }

  source.onopen = function() {

    console.log(arguments);

  };

  source.onerror = function () {

    console.log(arguments);

  };

  source.addEventListener('connections', log, false);

  source.addEventListener('requests', log, false);

  source.addEventListener('myevent', function(e) {

    data = JSON.parse(e.data);

    // .. do something..

  }, false);

  source.addEventListener('uptime', log, false);

  source.onmessage = function() {

    console.log(arguments);

  };

});

</script>

</body>

</html>

Publishing events

django_sse provides a convenience method to publish messages toRedisQueueView subclasses:

imoprt json

from django_sse.redisqueue import send_event

send_event('myevent', json.dumps(data))

Note that send_event allows only text values to be published. Taht's why we are serializing the data to json, and we unserialize it in the event handler withJSON.parse.

By default, django_sse publishes and listens to the redis channel see. If we want to separate messages per user, we can define the get_redis_channelmethod on the view:

class SSE(RedisQueueView):

    def get_redis_channel(self):

        return "sse_%s" % self.request.user.username

When we want to publish some event to a specific user, all we have to do is to specify the channel when calling send_event:

send_event('myevent', json.dumps(data), 'sse_%s' % user.username)

WebSocket

Now, suppose you want to notify user A when user B does some kind action.

You could still use SSEs, but every time the scenario happens, you'll end up with three connections: two long-running ones opened by A and B listening for SSEs, and a short one fired by B when POSTing his action.

Since you're already having long-running connections because you need to push events, you may just switch to WebSockets and save that POST.

Since WebSocket is not yet supported by Explorer, we'll have to use an abstraction layer, like socket.io or socks.js, that provide alternative transports of messages.

I choose to use socket.io mainly because of the gevent-socketio library, which integrates pretty easily with Django.

Using the socketio worker

In order to run gevent-socketio, we have to run gunicorn with a specialized worker class.

The GeventSocketIOWorker will take care of implementing the socket.io handshake and the new WebSocket Protocol (ws://)

In order to use GeventSocketIOWorker, I modified the worker_classparameter in the config file for unicorn:

#!python

from os import environ

from gevent import monkey

monkey.patch_all()

bind = "0.0.0.0:8000"

workers = 1

worker_class = "socketio.sgunicorn.GeventSocketIOWorker" # Note that we are now using gevent-socketio's worker

Note that using the socketio.sgunicorn.GeventSocketIOWorker is compatible with SSEs, so you could use this worker if you want both protocols running.

gevent-socketio allows you to define different Socket.io namespaces. This way you can implement different domain-specific logics. For example, you could implement a namespace for users' status (online, away, etc.) and a different chat messages.

Additionally, gevent-socketio ships with a couple of namespaces mixing for common situations, like for implementing separate chat rooms.

Let's create a namespace. Our namespace will provide separate chat-rooms, and will process events from our redis queue.

I had to override the emit_to_room method because I had the messages delivered more than once when I had more clients connected than the available workers.

myapp/sockets.py:

from socketio.namespace import BaseNamespace

from socketio.sdjango import namespace

from socketio.mixins import RoomsMixin

from myapp.utils import redis_connection

import json

@namespace('')

class MyNamespace(BaseNamespace, RoomsMixin):

    def listener(self, room):

        # ``redis_connection()`` is an utility function that returns a redis connection from a pool

        r = redis_connection().pubsub()

        r.subscribe('socketio_%s' % room)

        for m in r.listen():

            if m['type'] == 'message':

                data = json.loads(m['data'])

                self.process_event(data)

    def on_subscribe(self, *args):

        for channel in args:

            self.join(channel)

    def join(self, room):

        super(MyNamespace, self).join(room)

        self.spawn(self.listener, room)

        self.emit('joined', room)

    def on_myevent(self, room, *args):

        self.emit_to_room(room, 'myevent', *args)

    def emit_to_room(self, room, event, *args):

        """

        This is almost the same as ``.emit_to_room()`` on the parent class,

        but it sends events only over the current socket.

        This is to avoid a problem when there are more client than workers, and

        a single message can get delivered multiple times.

        """

        pkt = dict(type="event",

                   name=event,

                   args=args,

                   endpoint=self.ns_name)

        room_name = self._get_room_name(room)

        if 'rooms' not in self.socket.session:

            return

        if room_name in self.socket.session['rooms']:

            self.socket.send_packet(pkt)

Note that the join method we spawn a listener (and thus, a new redis connection) for every room we join. That's the way it's implemented in the chat example at the gevent-socketio repository.

If you're worried about having to spawn one process per client per channel, I've included an alternative subclass in the repo that restarts the listener when joining channel. The catch is that there will be a few milliseconds during which the user won't receive message.

I'm also using a pool to recycle Redis connection, The redis_connectionmethod creates a new redis object for our already existing connection pool:

utils.py:

from django.conf import settings

from redis import Redis

from redis import ConnectionPool as RedisConnectionPool

from redis.connection import Connection

WEBSOCKET_REDIS_BROKER_DEFAULT = {

    'HOST': 'localhost',

    'PORT': 6379,

    'DB': 0

}

CONNECTION_KWARGS = getattr(settings, 'WEBSOCKET_REDIS_BROKER', {})

class ConnectionPoolManager(object):

    """

    A singleton that contains and retrieves redis ``ConnectionPool``s according to the connection settings.

    """

    pools = {}

    @classmethod

    def key_for_kwargs(cls, kwargs):

        return ":".join([str(v) for v in kwargs.values()])

    @classmethod

    def connection_pool(cls, **kwargs):

        pool_key = cls.key_for_kwargs(kwargs)

        if pool_key in cls.pools:

            return cls.pools[pool_key]

        params = {

            'connection_class': Connection,

            'db': kwargs.get('DB', 0),

            'password': kwargs.get('PASSWORD', None),

            'host': kwargs.get('HOST', 'localhost'),

            'port': int(kwargs.get('PORT', 6379))

        }

        cls.pools[pool_key] = RedisConnectionPool(**params)

        return cls.pools[pool_key]

def redis_connection():

    """

    Returns a redis connection from one of our pools.

    """

    pool = ConnectionPoolManager.connection_pool(**CONNECTION_KWARGS)

    return Redis(connection_pool=pool)

For serving our namespaces, gevent-socketio gives us an autodiscovery feature similar to Django's admin:

urls.py:

from django.conf.urls import patterns, include, url

from myapp import views

import socketio.sdjango

socketio.sdjango.autodiscover()

urlpatterns = patterns('',

    url(r'^sse/$', views.SSE.as_view(), name='sse'),  # this URL is arbitrary.

    # socket.io uses the well-known URL `/socket.io/` for its protocol

    url(r"^socket\.io", include(socketio.sdjango.urls)),

    url(r'^$', views.HomePage.as_view(), name='homepage'),

)

On the client side, we need to include the socket.io javascript client, (available at https://github.com/LearnBoost/socket.io-client/).

By default, the client will try to use flashsockets under Internet Explorer (because Explorer doesn't support WebSocket).

The problem with flashsocket is that the Flash shipped with socketio-client makes a request for a policy file, and you'd need to set up a Flash policy server. So I decided to disable this transport and have IE use xhr-polling.

socket = io.connect('', {  // first argument is the namespace

  transports: ['websocket', 'xhr-multipart', 'xhr-polling', 'jsonp-polling']  // note ``flashsockets`` is missing

});

socket.on("myevent", function(e) {

  console.log("<myevent> event", arguments);

});

socket.on("message", function(e) {

  console.log("Message", e);

});

socket.on("joined", function(e) {

  console.log("joined", arguments);

});

socket.on("connect", function(e) {

  console.log("Connected", arguments);

  socket.emit('subscribe', 'default_room');

});

socket.on("disconnect", function(e) {

  console.log("Disconnected", arguments);

});

Publishing an event

All we have to do in order to emit an event to our client is pushing a message to the right redis channel.

utils.py:

# previous code here ...

import json

def emit_to_channel(channel, event, *data):

    r = redis_connection()

    args = [channel] + list(data)

    r.publish('socketio_%s' % channel, json.dumps({'name': event, 'args': args}))

Links & Acknowledgements

I would like to thank Jeff Triplett for the initial feedback on this article, Cody Soyland for his initial article about socket.io and gevent, Andrei Antoukh for accepting my patches for django_sse, and Jeffrey Gelens for accepting my patch for gevent-websocket.

If you want to read more, here's some links: