0.15

Subscriptions

Let's introduce a third type of operation. While queries offer a way to query a server once, subscriptions offer a way for the server to notify the client each time new data is available.

This is where the Subscription type is useful. It's similar to Query but as each subscription remains an open channel you can send anywhere from zero to millions of responses over its lifetime.

Because of their nature, subscriptions are only possible to implement in asynchronous servers that implement the WebSockets protocol. (If you are using uvicorn you need to pip install websockets otherwise you'll get Could not connect to websocket endpoint ws://localhost:8000/. Please check if the endpoint url is correct.)

WSGI-based servers (including Django) are synchronous in nature and unable to handle WebSockets which makes them incapable of implementing subscriptions.

If you wish to use subscriptions with Django, consider wrapping your Django application in a Django Channels container and using Ariadne as an ASGI server.

Note: Ariadne implements subscriptions-transport-ws protocol for GraphQL subscriptions.

Defining subscriptions

In schema definition subscriptions look similar to queries:

type Query {
    _unused: Boolean
}

type Subscription {
    counter: Int!
}

This example contains:

  • Query type with single unused field. GraphQL considers an empty type a syntax error and requires an API to always define a Query type.
    • For this example, we're focusing on Subscriptions so we define a bare bones Query type.
  • Subscription type with a single field, counter, that returns a number.

When defining subscriptions you can use all of the features of the schema such as arguments, input and output types.

Writing subscriptions

Subscriptions are more complex than queries as they require us to provide two functions for each field:

  • generator is a function that yields data we're going to send to the client. It has to implement the AsyncGenerator protocol.
  • resolver that tells the server how to send data to the client. This is similar to the resolvers we wrote earlier.

Make sure you understand how asynchronous generators work before attempting to use subscriptions.

The signatures are as follows:

async def counter_generator(obj, info):
    for i in range(5):
        await asyncio.sleep(1)
        yield i


def counter_resolver(count, info):
    return count + 1

Note that the resolver consumes the same type (in this case int) that the generator yields.

Each time our source yields a response, it's getting sent to our resolver. The above implementation counts from zero to four, each time waiting for one second before yielding a value.

The resolver increases each number by one before passing them to the client so the client sees the counter progress from one to five.

After the last value is yielded the generator returns, the server tells the client that no more data will be available, and the subscription is complete.

We can map these functions to subscription fields using the SubscriptionType class that extends ObjectType with support for source:

from ariadne import SubscriptionType
from . import counter_subscriptions

subscription = SubscriptionType()
subscription.set_field("counter", counter_subscriptions.counter_resolver)
subscription.set_source("counter", counter_subscriptions.counter_generator)

You can also use the source decorator:

@subscription.source("counter")
async def counter_generator(
    obj: Any, info: GraphQLResolveInfo
) -> AsyncGenerator[int, None]:
    ...

Publisher-consumer

Pubisher-consumer ("pub-sub") is a pattern in which parts of the system listen for ("subscribe to") events ("messages") from other parts of the system, usually reacting to them with very small delay.

To implement subscriptions, you will need to introduce a pub-sub solution to your stack. Multiple technologies are available here, starting with dedicated solutions like Apache Kafka, RabbitMQ and ending with data stores supporting subscribing to updates like Redis and PostgreSQL. Each of those solutions offers different features and trade offs, making them useful for different use-cases.

Only requirement by Ariadne is that technology has Python implementation that supports async subscriber.

Simple pub-sub setup with Broadcaster

Broadcaster is a simple pub-sub library that supports Redis, PostgreSQL and Apache Kafka as backends. It can be installed with pip:

pip install broadcaster

In our example we will use Redis server running on localhost at port 6379 for messaging. We instantiate Broadcaster with connection URL in our app:

broadcast = Broadcast("redis://localhost:6379")

We also need to run its connect and disconnect methods when our ASGI app starts or stops:

app = Starlette(on_startup=[broadcast.connect], on_shutdown=[broadcast.disconnect])

Publisher

We can publish our messages using the publish method:

await broadcast.publish(channel="chatroom", message="Hello world!")

Note: Channels are a way to group publishers and subscribers together. Your system may use single channel or multiple ones, each for different feature.

Where publishing code should live at? Simplest answer is at the same place that events occur that you would like your users to subscribe to. Here are few examples:

  • In GraphQL mutations: postComment mutation could publish event to notify other clients on same page that new commend was posted.
  • In task queues: process_video_file Celery task could publish event with current progress on processing uploaded video file.
  • In regular views: your JSON API or standard HTTP form view can send an event that contact form was sent to notify customer service members on-line.

Subscriber

Unlike publishers, which can go anywhere, subscribers in GraphQL API's have single dedicated place: subscriptions source:

@subscription.source("chat")
async def chat_generator(
    _: Any, info: GraphQLResolveInfo
) -> AsyncGenerator[str, None]:
    async with broadcast.subscribe(channel="chatroom") as subscriber:
        async for message in subscriber:
            yield message

In addition to that, generators can be used to filter which messages should and which shouldn't be sent further to the subscribers:

@subscription.source("chat")
async def chat_generator(
    _: Any, info: GraphQLResolveInfo
) -> AsyncGenerator[str, None]:
    swearwords = await load_swearwords()

    async with broadcast.subscribe(channel="chatroom") as subscriber:
        async for message in subscriber:
            if not contains_swearwords(message, swearwords):
                yield message

Simple chat example:

Here's example implementing simple GraphQL chat with mutation for sending messages and subscription for receiving them:

import json

from ariadne import MutationType, SubscriptionType, make_executable_schema
from ariadne.asgi import GraphQL
from broadcaster import Broadcast
from starlette.applications import Starlette


broadcast = Broadcast("memory://")


type_defs = """
  type Query {
    _unused: Boolean
  }

  type Message {
    sender: String
    message: String
  }

  type Mutation {
    send(sender: String!, message: String!): Boolean
  }

  type Subscription {
    message: Message
  }
"""


mutation = MutationType()


@mutation.field("send")
async def resolve_send(*_, **message):
    await broadcast.publish(channel="chatroom", message=json.dumps(message))
    return True


subscription = SubscriptionType()


@subscription.source("message")
async def source_message(_, info):
    async with broadcast.subscribe(channel="chatroom") as subscriber:
        async for event in subscriber:
            yield json.loads(event.message)


schema = make_executable_schema(type_defs, mutation, subscription)
graphql = GraphQL(schema=schema, debug=True)

app = Starlette(
    debug=True,
    on_startup=[broadcast.connect],
    on_shutdown=[broadcast.disconnect],
)

app.mount("/", graphql)

Connection params

Because subscriptions in GraphQL are done over the websockets, you can't use custom HTTP headers to pass additional data from client to server. This makes it impossible to use Authorization header for authentication within subscriptions.

To work around this limitation, websocket clients include this data in initial message sent to the server as part of connection negotiation.

Using on_connect to access connection's parameters

To access connection parameters, custom function needs to be implemented and passed to Ariadne's on_connect option:

def on_connect(websocket, params: Any):
    ...


graphql = GraphQL(schema, on_connect=on_connect)

This function is called exactly once: at the time when websocket connection is opened by the client. It's always called with two arguments: a starlette.websockets.WebSocket instance and a payload. It can be synchronous or asynchronous.

Please note that because params value is set by the client there are no guarantees on what type and concents this value will be. Due care needs to be taken here:

def on_connect(websocket, params: Any):
    if not isinstance(params, dict):
        return

    token = params.get("token")
    if token:
        ...

In order to make params available to the resolvers, they need to be passed through the WebSocket.scope dict to context factory:

def on_connect(websocket, params: Any):
    if not isinstance(params, dict):
        websocket.scope["connection_params"] = {}
        return

    # websocket.scope is a dict acting as a "bag"
    # stores data for the duration of connection
    websocket.scope["connection_params"] = {
        "token": params.get("token"),
    }


def context_value(request):
    context = {}

    if request.scope["type"] == "websocket":
        # request is an instance of WebSocket
        context.update(request.scope["connection_params"])
    else:
        context["token"] = request.META.get("authorization")

    return context


graphql = GraphQL(schema, context_value=context_value, on_connect=on_connect)

on_connect vs context_value

There's important difference between on_connect and context_value:

on_connect is called once, at the time of websocket connection negotiation between client and GraphQL server.

context_value is called every time new subscription query is made by the client.

If your client has two separate UI components (eg. notification bell on the navbar and list of on-line users), and those components do GraphQL subscribe queries, context_value will be ran for each of those separately while on_connect will only be ran once.

Note: This behavior is true for most popular GraphQL client implementations (gql and Apollo-Client) but may not be true for some libraries.

This can have implications for application performance. It may be preferable to cache data on websocket.scope instead of info.context to avoid repeated database reads for multiple subscriptions accessing same data. Or pre-load user object in on_connect.

Refusing websocket connection

To refuse websocket connection from client, you can raise ariadne.asgi.WebSocketConnectionError from on_connect:

def on_connect(websocket, params: Any):
    if not isinstance(params, dict):
        raise WebSocketConnectionError("Invalid payload")

    token = params.get("token")
    if not token:
        raise WebSocketConnectionError("Missing auth")

If you have control on client implementation as well, you can pass custom error payload instead of string:

def on_connect(websocket, params: Any):
    if not isinstance(params, dict):
        raise WebSocketConnectionError({"message": "Invalid payload", "code": "invalid_payload"})

    token = params.get("token")
    if not token:
        raise WebSocketConnectionError({"message": "Missing auth", "code": "auth"})

on_operation and on_complete

Warning: This feature is considered experimental. It was implemented for feature parity with older version of Apollo Server. Its final shape (or presence in future Ariadne releases) is snot decided yet. Generally you should try using on_connect and on_disconnect first before using those features.

on_operation and on_complete options allow you to run extra python code when client subscribes or unsubscribes from Subscription field within same WebSocket connection:

def on_operation(websocket, operation: Operation):
    ...


def on_complete(websocket, operation: Operation):
    ...


graphql = GraphQL(schema, on_operation=on_operation, on_complete=on_complete)

First argument for those functions is WebSocket instance and second one is Operation dataclass storing data about current subscription:

@dataclass
class Operation:
    id: str
    name: Optional[str]
    generator: AsyncGenerator

on_disconnect

on_disconnect option can be set to callable function taking single argument, WebSocket instance, that should be ran after Ariadne closes the websocket connection:

def on_connect(websocket, params: Any):
    if not isinstance(params, dict):
        websocket.scope["connection_params"] = {}
        return

    chat_user = get_user_from_ws(params)
    chat_user.set_online()
    websocket.scope["chat_user"] = chat_user


def on_disconnect(websocket):
    chat_user = websocket.scope.get("chat_user")
    if chat_user:
        chat_user.set_offline()


graphql = GraphQL(schema, on_connect=on_connect, on_disconnect=on_disconnect)
← Interface typesFile uploads →