Django Channels based WebSocket GraphQL server with Graphene-like subscriptions

Django Channels based WebSocket GraphQL server with Graphene-like subscriptions

Features

WebSocket-based GraphQL server implemented on the Django Channels v3.
WebSocket protocol is compatible with Apollo GraphQL client.
Graphene-like subscriptions.
All GraphQL requests are processed concurrently (in parallel).
Subscription notifications delivered in the order they were issued.
Optional subscription activation message can be sent to a client. This is useful to avoid race conditions on the client side. Consider the case when client subscribes to some subscription and immediately invokes a mutations which triggers this subscription. In such case the subscription notification can be lost, cause these subscription and mutation requests are processed concurrently. To avoid this client shall wait for the subscription activation message before sending such mutation request.
Customizable notification strategies:
- A subscription can be put to one or many subscription groups. This allows to granularly notify only selected clients, or, looking from the client's perspective - to subscribe to some selected source of events. For example, imaginary subscription "OnNewMessage" may accept argument "user" so subscription will only trigger on new messages from the selected user.
- Notification can be suppressed in the subscription resolver method publish. For example, this is useful to avoid sending self-notifications.
All GraphQL "resolvers" run in the main eventloop. Asynchronous "resolvers" able to execute blocking calls with asyncio.to_thread or channels.db.database_sync_to_async wrappers.
Resolvers (including subscription's subscribe & publish) can be represented both as synchronous or asynchronous (async def) methods.
Subscription notifications can be sent from both synchronous and asynchronous contexts. Just call MySubscription.broadcast() or await MySubscription.broadcast() depending on the context.
Clients for the GraphQL WebSocket server:
- AIOHTTP-based client.
- Client for unit test based on the Channels testing communicator.
Requires Python 3.8 and newer. Tests run on 3.8, 3.9, 3.10.
Works on Linux, macOS, and Windows.

Installation

pip install django-channels-graphql-ws

Getting started

Create a GraphQL schema using Graphene. Note the MySubscription class.

import channels_graphql_ws
import graphene

class MySubscription(channels_graphql_ws.Subscription):
    """Simple GraphQL subscription."""

    # Leave only latest 64 messages in the server queue.
    notification_queue_limit = 64

    # Subscription payload.
    event = graphene.String()

    class Arguments:
        """That is how subscription arguments are defined."""
        arg1 = graphene.String()
        arg2 = graphene.String()

    @staticmethod
    def subscribe(root, info, arg1, arg2):
        """Called when user subscribes."""

        # Return the list of subscription group names.
        return ["group42"]

    @staticmethod
    def publish(payload, info, arg1, arg2):
        """Called to notify the client."""

        # Here `payload` contains the `payload` from the `broadcast()`
        # invocation (see below). You can return `None` if you wish to
        # suppress the notification to a particular client. For example,
        # this allows to avoid notifications for the actions made by
        # this particular client.

        return MySubscription(event="Something has happened!")

class Query(graphene.ObjectType):
    """Root GraphQL query."""
    # Graphene requires at least one field to be present. Check
    # Graphene docs to see how to define queries.
    value = graphene.String()
    async def resolve_value(self):
        return "test"

class Mutation(graphene.ObjectType):
    """Root GraphQL mutation."""
    # Check Graphene docs to see how to define mutations.
    pass

class Subscription(graphene.ObjectType):
    """Root GraphQL subscription."""
    my_subscription = MySubscription.Field()

graphql_schema = graphene.Schema(
    query=Query,
    mutation=Mutation,
    subscription=Subscription,
)

Make your own WebSocket consumer subclass and set the schema it serves:

class MyGraphqlWsConsumer(channels_graphql_ws.GraphqlWsConsumer):
    """Channels WebSocket consumer which provides GraphQL API."""
    schema = graphql_schema

    # Uncomment to send keepalive message every 42 seconds.
    # send_keepalive_every = 42

    # Uncomment to process requests sequentially (useful for tests).
    # strict_ordering = True

    async def on_connect(self, payload):
        """New client connection handler."""
        # You can `raise` from here to reject the connection.
        print("New client connected!")

Setup Django Channels routing:

application = channels.routing.ProtocolTypeRouter({
    "websocket": channels.routing.URLRouter([
        django.urls.path("graphql/", MyGraphqlWsConsumer.as_asgi()),
    ])
})

Notify^﹡ clients when some event happens using the broadcast() or broadcast_sync() method from the OS thread where there is no running event loop:

MySubscription.broadcast(
    # Subscription group to notify clients in.
    group="group42",
    # Dict delivered to the `publish` method.
    payload={},
)

Notify^﹡ clients in a coroutine function with async broadcast() or broadcast_async() method:

await MySubscription.broadcast(
    # Subscription group to notify clients in.
    group="group42",
    # Dict delivered to the `publish` method.
    payload={},
)

﹡) In case you are testing your client code by notifying it from the Django Shell, you have to setup a channel layer in order for the two instance of your application. The same applies in production with workers.

You should prefer async resolvers and async middleware over sync ones. Async versions will result in faster code execution. To do DB operations you can use Django 4 asynchronous queries.

Example

You can find simple usage example in the example directory.

Run:

cd example/
# Initialize database.
./manage.py migrate
# Create "user" with password "user".
./manage.py createsuperuser
# Run development server.
./manage.py runserver

Play with the API though the GraphiQL browser at http://127.0.0.1:8000.

You can start with the following GraphQL requests:

# Check there are no messages.
query read { history(chatroom: "kittens") { chatroom text sender }}

# Send a message as Anonymous.
mutation send { sendChatMessage(chatroom: "kittens", text: "Hi all!"){ ok }}

# Check there is a message from `Anonymous`.
query read { history(chatroom: "kittens") { text sender } }

# Login as `user`.
mutation send { login(username: "user", password: "pass") { ok } }

# Send a message as a `user`.
mutation send { sendChatMessage(chatroom: "kittens", text: "It is me!"){ ok }}

# Check there is a message from both `Anonymous` and from `user`.
query read { history(chatroom: "kittens") { text sender } }

# Subscribe, do this from a separate browser tab, it waits for events.
subscription s { onNewChatMessage(chatroom: "kittens") { text sender }}

# Send something again to check subscription triggers.
mutation send { sendChatMessage(chatroom: "kittens", text: "Something ;-)!"){ ok }}

Details

The channels_graphql_ws module provides the following key classes:

GraphqlWsConsumer: Django Channels WebSocket consumer which maintains WebSocket connection with the client.
Subscription: Subclass this to define GraphQL subscription. Very similar to defining mutations with Graphene. (The class itself is a "creative" copy of the Graphene Mutation class.)
GraphqlWsClient: A client for the GraphQL backend. Executes strings with queries and receives subscription notifications.
GraphqlWsTransport: WebSocket transport interface for the client.
GraphqlWsTransportAiohttp: WebSocket transport implemented on the AIOHTTP library.

For details check the source code which is thoroughly commented. The docstrings of classes are especially useful.

Since the WebSocket handling is based on the Django Channels and subscriptions are implemented in the Graphene-like style it is recommended to have a look the documentation of these great projects:

The implemented WebSocket-based protocol was taken from the library subscription-transport-ws which is used by the Apollo GraphQL. Check the protocol description for details.

Automatic Django model serialization

The Subscription.broadcast uses Channels groups to deliver a message to the Subscription's publish method. ASGI specification clearly states what can be sent over a channel, and Django models are not in the list. Since it is common to notify clients about Django models changes we manually serialize the payload using MessagePack and hack the process to automatically serialize Django models following the the Django's guide Serializing Django objects.

Execution

Different requests from different WebSocket client are processed asynchronously.
By default different requests (WebSocket messages) from a single client are processed concurrently in an event loop. So there is no guarantee that requests will be processed in the same order the client sent these requests. Actually, with HTTP we have this behavior for decades.
It is possible to serialize message processing by setting strict_ordering to True. But note, this disables parallel requests execution - in other words, the server will not start processing a new request from the client until it finishes the current one. See comments in the class GraphqlWsConsumer. This mode is here primarily for testing.
All subscription notifications are delivered in the order they were issued.
Each request (WebSocket message) processing starts in the main thread. The request's parsing and validation is offloaded into the thread pool. Resolver calls made from the main thread. And for each resolver it checks whether the resolver is awaitable and await it if so.

Context and scope

The context object (info.context in resolvers) is a SimpleNamespace instance useful to transfer extra data between GraphQL resolvers. The lifetime of info.context corresponds to the lifetime of GraphQL request, so it does not persist content between different queries/mutations/subscriptions. It also contains some useful extras:

graphql_operation_id: The GraphQL operation id came from the client.
graphql_operation_name: The name of GraphQL operation.
channels_scope: Channels scope. Contrary to the info.context, the Channels scope corresponds to the WebSocket connection not to the GraphQL operation/request.
channel_name: WebSocket channel name.

Authentication

To enable authentication it is typically enough to wrap your ASGI application into the channels.auth.AuthMiddlewareStack:

application = channels.routing.ProtocolTypeRouter({
    "websocket": channels.auth.AuthMiddlewareStack(
        channels.routing.URLRouter([
            django.urls.path("graphql/", MyGraphqlWsConsumer),
        ])
    ),
})

This gives you a Django user info.context.channels_scope["user"] in all the resolvers. To authenticate user you can create a Login mutation like the following:

class Login(graphene.Mutation, name="LoginPayload"):
    """Login mutation."""

    ok = graphene.Boolean(required=True)

    class Arguments:
        """Login request arguments."""

        username = graphene.String(required=True)
        password = graphene.String(required=True)

    def mutate(self, info, username, password):
        """Login request."""

        # Ask Django to authenticate user.
        user = django.contrib.auth.authenticate(username=username, password=password)
        if user is None:
            return Login(ok=False)

        # Use Channels to login, in other words to put proper data to
        # the session stored in the scope.
        asgiref.sync.async_to_sync(channels.auth.login)(info.context.channels_scope, user)
        # Save the session,cause `channels.auth.login` does not do this.
        info.context.session.save()

        return Login(ok=True)

The authentication is based on the Channels authentication mechanisms. Check the Channels documentation. Also take a look at the example in the example directory.

The Python client

There is the GraphqlWsClient which implements GraphQL client working over the WebSockets. The client needs a transport instance which communicates with the server. Transport is an implementation of the GraphqlWsTransport interface (class must be derived from it). There is the GraphqlWsTransportAiohttp which implements the transport on the AIOHTTP library. Here is an example:

transport = channels_graphql_ws.GraphqlWsTransportAiohttp(
    "ws://backend.endpoint/graphql/", cookies={"sessionid": session_id}
)
client = channels_graphql_ws.GraphqlWsClient(transport)
await client.connect_and_init()
result = await client.execute("query { users { id login email name } }")
users = result["data"]
await client.finalize()

See the GraphqlWsClient class docstring for the details.

The GraphiQL client

The GraphiQL provided by Graphene doesn't connect to your GraphQL endpoint via WebSocket; instead you should use a modified GraphiQL template under graphene/graphiql.html which will take precedence over the one of Graphene. One such modified GraphiQL is provided in the example directory.

Testing

To test GraphQL WebSocket API read the appropriate page in the Channels documentation.

In order to simplify unit testing there is a GraphqlWsTransport implementation based on the Django Channels testing communicator: channels_graphql_ws.testing.GraphqlWsTransport. Check its docstring and take a look at the tests to see how to use it.

Subscription activation confirmation

The original Apollo's protocol does not allow client to know when a subscription activates. This inevitably leads to the race conditions on the client side. Sometimes it is not that crucial, but there are cases when this leads to serious issues. Here is the discussion in the subscriptions-transport-ws tracker.

To solve this problem, there is the GraphqlWsConsumer setting confirm_subscriptions which when set to True will make the consumer issue an additional data message which confirms the subscription activation. Please note, you have to modify the client's code to make it consume this message, otherwise it will be mistakenly considered as the first subscription notification.

To customize the confirmation message itself set the GraphqlWsConsumer setting subscription_confirmation_message. It must be a dictionary with two keys "data" and "errors". By default it is set to {"data": None, "errors": None}.

GraphQL middleware

It is possible to inject middleware into the GraphQL operation processing. For that define middleware setting of your GraphqlWsConsumer subclass, like this:

async def threadpool_for_sync_resolvers(next_middleware, root, info, *args, **kwds):
    """Offload synchronous resolvers to the threadpool.

    This middleware should always be the last in the middlewares calls
    stack and the closest to the real resolver. If this middleware is
    not the last it will check the next middleware to call instead of
    real resolver. The first middleware in the middlewares list will be
    the closest to the resolver.
    """
    # Invoke next middleware.
    if asyncio.iscoroutinefunction(next_middleware):
        result = await next_middleware(root, info, *args, **kwds)
    else:
        result = await asyncio.to_thread(next_middleware, root, info, *args, **kwds)
    return result

class MyGraphqlWsConsumer(channels_graphql_ws.GraphqlWsConsumer):
    ...
    middleware = [threadpool_for_sync_resolvers]

It is recommended to write asynchronous middlewares. But synchronous middlewares are also supported:

def my_middleware(next_middleware, root, info, *args, **kwds):
    """My custom GraphQL middleware."""
    # Invoke next middleware.
    return next_middleware(root, info, *args, **kwds)

class MyGraphqlWsConsumer(channels_graphql_ws.GraphqlWsConsumer):
    ...
    middleware = [my_middleware]

For more information about GraphQL middleware please take a look at the relevant section in the Graphene documentation.

Alternatives

There is a Tomáš Ehrlich GitHubGist GraphQL Subscription with django-channels which this implementation was initially based on.

There is a promising GraphQL WS library by the Graphene authors. In particular this pull request gives a hope that there will be native Graphene implementation of the WebSocket transport with subscriptions one day.

Development

Bootstrap

A reminder of how to setup an environment for the development.

Install PyEnv to be able to work with many Python versions at once PyEnv→Installation.
Install Python versions needed. The command should be executed in the project's directory:
```
$ pyenv local | xargs -L1 pyenv install
```
Check that pyenv works correctly. The command:
```
$ pyenv versions
```
should show python versions enlisted in .python-version. If everything is set up correctly pyenv will switch version of python when you enter and leave the project's directory. Inside the directory pyenv which python should show you a python installed in pyenv, outside the dir it should be the system python.
Install Poetry to the system Python.
```
$ curl -sSL https://install.python-poetry.org | python3 -
```
It is important to install Poetry into the system Python, NOT in your virtual environment. For details see Poetry docs: https://python-poetry.org/docs/#installation
Create local virtualenv in .venv, install all project dependencies (from pyproject.toml) except the project itself.
```
$ poetry install --no-root
```
Activate virtualenv There are options:
- With Poetry:
```
$ poetry shell
```
- Manually:
```
$ source .venv/bin/activate
```
- With VS Code: Choose .venv with "Python: Select interpreter" and reopen the terminal.
Upgrade Pip:
```
$ pip install --upgrade pip
```

Use:

Where to start reading the code

The code is inherently complex because it glues two rather different libraries/frameworks Channels and Graphene. You might need some time to dive into. Here are some quick insights to help you to get on track.

The main classes are GraphqlWsConsumer and Subscription. The former one is a Channels consumer which instantiates each time a WebSocket connection establishes. User (of the library) subclasses it and tunes settings in the successor class. The latter is from the Graphene world. Both classes are tightly coupled. When client subscribes an instance of GraphqlWsConsumer subclass holding the WebSocket connection passes to the Subscription.

To better dive in it is useful to understand in general terms how regular request are handled. When server receives JSON from the client, the GraphqlWsConsumer.receive_json method is called by Channels routines. Then the request passes to the _on_gql_start method which handles GraphQL message "START". Most magic happens there.

Running tests

A reminder of how to run tests.

Run all tests on all supported Python versions:
```
$ tox
```
Run all tests on a single Python version, e.g on Python 3.8:
```
$ tox -e py38
```

Example of running a single test:

$ tox -e py310 -- tests/test_basic.py::test_main_usecase

Running on currently active Python directly with Pytest:
```
$ poetry run pytest
```

Making release

A reminder of how to make and publish a new release.

Merge all changes to the master branch and switch to it.
Update version: poetry version minor.
Update CHANGELOG.md.
Update README.md (if needed).
Commit changes made above.
Git tag: git tag vX.X.X && git push --tags.
Publish release to PyPI: poetry publish --build.
Update release notes on GitHub.

Contributing

This project is developed and maintained by DATADVANCE LLC. Please submit an issue if you have any questions or want to suggest an improvement.

Acknowledgements

This work is supported by the Russian Foundation for Basic Research (project No. 15-29-07043).

datadvance / DjangoChannelsGraphqlWs