channels_rabbitmq

A Django Channels channel layer that uses RabbitMQ as its backing store.

Does not support Worker and Background Tasks. (See Rationale and use await get_channel_layer().carehare_connection to send to job queues.)

Works with Python 3.8 or 3.9.

Installation

pip install channels_rabbitmq

Usage

Then set up the channel layer in your Django settings file like so:

CHANNEL_LAYERS = {
    "default": {
        "BACKEND": "channels_rabbitmq.core.RabbitmqChannelLayer",
        "CONFIG": {
            "host": "amqp://guest:guest@127.0.0.1/asgi",
            # "ssl_context": ... (optional)
        },
    },
}

Possible options for CONFIG are listed below.

host

URL of the server to connect to, adhering to RabbitMQ spec. To connect to a RabbitMQ cluster, use a DNS server to resolve a hostname to multiple IP addresses. channels_rabbitmq will automatically reconnect if at least one of them is reachable in case of a disconnection.

expiry

Minimum number of seconds a message should wait in a RabbitMQ queue, before it may be silently dropped.

Defaults to 60. You generally shouldn't need to change this, but you may want to turn it down if you have peaky traffic you wish to drop, or up if you have peaky traffic you want to backlog until you get to it.

local_capacity

Number of incoming messages queued in memory. Defaults to 100. (A message sent to a group with two channels counts as one message.) When local_capacity messages are queued, the message backlog will grow on RabbitMQ.

(This controls the prefetch_count on the RabbitMQ queue.)

local_expiry

Minimum number of seconds a message received from RabbitMQ must be held in memory waiting for receive(), before it may be dropped. Defaults to expiry.

A warning will be logged when a message expires locally. The warning can indicate that a channel has more messages than it can handle; or that messages are being sent to a channel that does not exist. (Perhaps a missing channel was implied by group_add(), and a matching group_discard() was never called.)

If local_expiry < expiry, then you can end up ignoring (and logging) messages locally while they still exist in the RabbitMQ queue. These messages will be acked, so RabbitMQ will behave as though they were delivered.

remote_capacity

Number of messages stored on RabbitMQ for each client. Defaults to 100. (A message sent to a group with three channels on two distinct clients counts as two messages.) When remote_capacity messages are queued in RabbitMQ, the channel will refuse new messages. Calls from any client to send() or group_send() to the at-capacity client will raise ChannelFull.

ssl_context

An SSL context. Changes the default host port to 5671 (instead of 5672).

For instance, to connect to an TLS RabbitMQ service that will verify your client:

import ssl
ssl_context = ssl.create_default_context(
    cafile=str(Path(__file__).parent.parent / 'ssl' / 'server.cert'),
)
ssl_context.load_cert_chain(
    certfile=str(Path(__file__).parent.parent / 'ssl' / 'client.certchain'),
    keyfile=str(Path(__file__).parent.parent / 'ssl' / 'client.key'),
)
CHANNEL_LAYERS['default']['CONFIG']['ssl_context'] = ssl_context

By default, there is no SSL context; all messages (and passwords) are are transmitted in cleartext.

groups_exchange

Global direct exchange name used by channels to exchange group messages. Defaults to "groups". See also Design decisions.

Accessing Carehare

We use carehare for its thorough handling of errors.

Django Channels' specification does not account for "connecting" and "disconnecting". This layer does its best by constantly reconnecting, forever.

Call await get_channel_layer().carehare_connection to access an open Carehare connection. This lets you use job queues without "Worker and Background Tasks". Like this:

# raise asyncio.CancelledError on failure
connection = await get_channel_layer().carehare_connection

# raise carehare.ConnectionClosed or carehare.ChannelClosed on error
await connection.publish(b"task", routing_key="job_queue")

(The Carehare documentation explains how to build workers.)

A note on errors: a "connected" connection isn't guaranteed to stay connected throughout every publish. It was merely connected at some point in the past. When a disconnect occurs, all pending operations on that connection will raise carehare.ConnectionClosed. This channel layer will log the error, and get_channel_layer().carehare_connection will point to a new Future. (This error+reconnect is guaranteed to happen in production.)

Publish messages from Celery

Many Celery users want to send messages to websockets users.

This is doable, though not intuitive. Don't use any Django Channels code: Channels layers depend on long-running connections, and Celery bans those. Same goes for carehare: don't use it from Celery.

Instead, from a Celery-worker @task you can send messages to your Django-Channels consumers using Celery's RabbitMQ connection:

from typing import Any, Dict

import msgpack

def publish_message_to_group(message: Dict[str, Any], group: str) -> None:
    with current_app.producer_pool.acquire(block=True) as producer:
        producer.publish(
            msgpack.packb({
              "__asgi_group__": group,
              **message,
            }),
            exchange="groups",  # groups_exchange
            content_encoding="binary",
            routing_key=group,
            retry=False,  # Channel Layer at-most once semantics
        )

To call it, from a Celery-worker @task...:

publish_message_to_group({ "type": "chat.message", "text": "hi" }, "a-group")

... and a Django-Channels consumer like this will receive it:

class WebsocketConnectionConsumer(AsyncWebsocketConsumer):
    async def connect(self):
        await self.channel_layer.group_add("a-group", self.channel_name)

    async def disconnect(self):
        await self.channel_layer.group_discard("a-group", self.channel_name)

    async def chat_message(self, event):
        assert event["text"] == "hi"

Alternatively, write your workers asynchronously, directly in carehare. It's more lightweight and faster than Celery, and the error handling is simpler.

Design decisions

To scale enormously, this layer only creates one RabbitMQ queue per instance. That means one web server gets one RabbitMQ queue, no matter how many websocket connections are open. For each message being sent, the client-side layer determines the RabbitMQ queue name and uses it as the routing key.

Groups are implemented using a single, global RabbitMQ direct exchange called "groups" by default. To send a message to a group, the layer sends the message to the "groups" exchange with the group name as the routing key. The client binds and unbinds during group_add() and group_remove() to ensure messages for any of its groups will reach it. See also the groups_exchange option.

RabbitMQ queues are exclusive: when a client disconnects (through close or crash), RabbitMQ will delete the queue and unbind the groups.

Once a connection has been created, it pollutes the event loop so that async_to_sync() will destroy the connection if it was created within async_to_sync(). Each connection starts a background async loop that pulls messages from RabbitMQ and routes them to receiver queues; each receive() queries receiver queues. Empty queues with no connections are deleted.

Deviations from the Channel Layer Specification

The Channel Layer Specification bends to Redis-related restrictions. RabbitMQ cannot emulate Redis. Here are the differences:

• No ``flush`` extension: To flush all state, simply disconnect all clients. (RabbitMQ won't allow one client to delete another client's data structures.)

• No ``group_expiry`` option: The group_expiry option recovers when a group_add() has no matching group_discard(). But the "group membership expiry" logic has a fatal flaw: it disconnects legitimate members. channels_rabbitmq addresses each root problem instead:

  • Web-server crash: RabbitMQ wipes all state related to a web server when the web server disconnects. There's no problem here for group_expiry to solve.
  • Programming errors: You may err and call group_add() without eventually calling group_discard(). Redis can't detect this programming error (because it can't detect web-server crashes). RabbitMQ can. The local_expiry option keeps your site running after you erroneously miss a group_discard(). The channel layer warns when discarding expired messages. Monitor your server logs to detect your errors.

• No "normal channels": Normal channels are job queues. In most projects, "normal channel" readers are worker processes, ideally divorced from Websockets and Django.

  If you want an async, RabbitMQ-based job queue, investigate carehare.

  If you're using Celery with the same RabbitMQ server, you can publish messages from Celery, too.

Dependencies

You'll need Python 3.8+ and a RabbitMQ server.

If you have Docker, here's how to start a development server:

ssl/prepare-certs.sh  # Create SSL certificates used in tests
docker run --rm -it \
     -p 5671:5671 \
     -p 5672:5672 \
     -p 15672:15672 \
     -v "/$(pwd)"/ssl:/ssl \
     -e RABBITMQ_SSL_CACERTFILE=/ssl/ca.cert \
     -e RABBITMQ_SSL_CERTFILE=/ssl/server.cert \
     -e RABBITMQ_SSL_KEYFILE=/ssl/server.key \
     -e RABBITMQ_SSL_VERIFY=verify_peer \
     -e RABBITMQ_SSL_FAIL_IF_NO_PEER_CERT=true \
     rabbitmq:3.7.8-management-alpine

You can access the RabbitMQ management interface at http://localhost:15672.

Contributing

To add features and fix bugs

First, start a development RabbitMQ server:

ssl/prepare-certs.sh  # Create SSL certificates used in tests
docker run --rm -it \
     -p 5671:5671 \
     -p 5672:5672 \
     -p 15672:15672 \
     -v "/$(pwd)"/ssl:/ssl \
     -e RABBITMQ_SSL_CACERTFILE=/ssl/ca.cert \
     -e RABBITMQ_SSL_CERTFILE=/ssl/server.cert \
     -e RABBITMQ_SSL_KEYFILE=/ssl/server.key \
     -e RABBITMQ_SSL_VERIFY=verify_peer \
     -e RABBITMQ_SSL_FAIL_IF_NO_PEER_CERT=true \
     rabbitmq:3.8.11-management-alpine

Now take on the development cycle:

1. tox # to ensure tests pass.
2. Write new tests in tests/ and make sure they fail.
3. Write new code in channels_rabbitmq/ to make the tests pass.
4. Submit a pull request.

To deploy

Use semver.

1. git push and make sure Travis tests all pass.
2. git tag vX.X.X
3. git push --tags

TravisCI will push to PyPi.