Subscriptions in Apollo Server

Subscriptions are long-lasting GraphQL read operations that can update their result whenever a particular server-side event occurs. Most commonly, updated results are pushed from the server to subscribing clients. For example, a chat application's server might use a subscription to push newly received messages to all clients in a particular chat room.

Because subscription updates are usually pushed by the server (instead of polled by the client), they generally use the WebSocket protocol instead of HTTP.

Important: Compared to queries and mutations, subscriptions are significantly more complex to implement. Before you begin, confirm that your use case requires subscriptions.

Schema definition

Your schema's Subscription type defines top-level fields that clients can subscribe to:

1type Subscription {
2  postCreated: Post
3}

The postCreated field will update its value whenever a new Post is created on the backend, thus pushing the Post to subscribing clients.

Clients can subscribe to the postCreated field with a GraphQL string, like this:

1subscription PostFeed {
2  postCreated {
3    author
4    comment
5  }
6}

Each subscription operation can subscribe to only one field of the Subscription type.

Enabling subscriptions

Subscriptions are not supported by Apollo Server's startStandaloneServer function. To enable subscriptions, you must first swap to using the expressMiddleware function (or any other Apollo Server integration package that supports subscriptions).

The following steps assume you've already swapped to expressMiddleware.

To run both an Express app and a separate WebSocket server for subscriptions, we'll create an http.Server instance that effectively wraps the two and becomes our new listener.

1. Install graphql-ws, ws, and @graphql-tools/schema:

   Bash

   copy

   1npm install graphql-ws ws @graphql-tools/schema

2. Add the following imports to the file where you initialize your ApolloServer instance (we'll use these in later steps):

   TypeScript

   index.ts

   copy

   1import { createServer } from 'http';
   2import { ApolloServerPluginDrainHttpServer } from '@apollo/server/plugin/drainHttpServer';
   3import { makeExecutableSchema } from '@graphql-tools/schema';
   4import { WebSocketServer } from 'ws';
   5import { useServer } from 'graphql-ws/lib/use/ws';

3. Next, in order to set up both the HTTP and subscription servers, we need to first create an http.Server. Do this by passing your Express app to the createServer function, which we imported from the http module:

   TypeScript

   index.ts

   copy

   1// This `app` is the returned value from `express()`.
   2const httpServer = createServer(app);

4. Create an instance of GraphQLSchema (if you haven't already).

   If you already pass the schema option to the ApolloServer constructor (instead of typeDefs and resolvers), you can skip this step.

   The subscription server (which we'll instantiate next) doesn't take typeDefs and resolvers options. Instead, it takes an executable GraphQLSchema. We can pass this schema object to both the subscription server and ApolloServer. This way, we make sure that the same schema is being used in both places.

   TypeScript

   index.ts

   copy

   1const schema = makeExecutableSchema({ typeDefs, resolvers });
   2// ...
   3const server = new ApolloServer({
   4  schema,
   5});

5. Create a WebSocketServer to use as your subscription server.

   TypeScript

   index.ts

   copy

   1// Creating the WebSocket server
   2const wsServer = new WebSocketServer({
   3  // This is the `httpServer` we created in a previous step.
   4  server: httpServer,
   5  // Pass a different path here if app.use
   6  // serves expressMiddleware at a different path
   7  path: '/subscriptions',
   8});
   9
   10// Hand in the schema we just created and have the
   11// WebSocketServer start listening.
   12const serverCleanup = useServer({ schema }, wsServer);

6. Add plugins to your ApolloServer constructor to shutdown both the HTTP server and the WebSocketServer:

   TypeScript

   index.ts

   copy

   1const server = new ApolloServer({
   2  schema,
   3  plugins: [
   4    // Proper shutdown for the HTTP server.
   5    ApolloServerPluginDrainHttpServer({ httpServer }),
   6
   7    // Proper shutdown for the WebSocket server.
   8    {
   9      async serverWillStart() {
   10        return {
   11          async drainServer() {
   12            await serverCleanup.dispose();
   13          },
   14        };
   15      },
   16    },
   17  ],
   18});

7. Finally, ensure you are listening to your httpServer.

   Most Express applications call app.listen(...), but for your setup change this to httpServer.listen(...) using the same arguments. This way, the server starts listening on the HTTP and WebSocket transports simultaneously.

A completed example of setting up subscriptions is shown below:

1import { ApolloServer } from '@apollo/server';
2import { ApolloServerPluginDrainHttpServer } from '@apollo/server/plugin/drainHttpServer';
3import { expressMiddleware } from '@as-integrations/express5';
4import { createServer } from 'http';
5import express from 'express';
6import { makeExecutableSchema } from '@graphql-tools/schema';
7import { WebSocketServer } from 'ws';
8import { useServer } from 'graphql-ws/lib/use/ws';
9import cors from 'cors';
10import resolvers from './resolvers';
11import typeDefs from './typeDefs';
12
13// Create the schema, which will be used separately by ApolloServer and
14// the WebSocket server.
15const schema = makeExecutableSchema({ typeDefs, resolvers });
16
17// Create an Express app and HTTP server; we will attach both the WebSocket
18// server and the ApolloServer to this HTTP server.
19const app = express();
20const httpServer = createServer(app);
21
22// Create our WebSocket server using the HTTP server we just set up.
23const wsServer = new WebSocketServer({
24  server: httpServer,
25  path: '/subscriptions',
26});
27// Save the returned server's info so we can shutdown this server later
28const serverCleanup = useServer({ schema }, wsServer);
29
30// Set up ApolloServer.
31const server = new ApolloServer({
32  schema,
33  plugins: [
34    // Proper shutdown for the HTTP server.
35    ApolloServerPluginDrainHttpServer({ httpServer }),
36
37    // Proper shutdown for the WebSocket server.
38    {
39      async serverWillStart() {
40        return {
41          async drainServer() {
42            await serverCleanup.dispose();
43          },
44        };
45      },
46    },
47  ],
48});
49
50await server.start();
51app.use(
52  '/graphql',
53  cors<cors.CorsRequest>(),
54  express.json(),
55  expressMiddleware(server),
56);
57
58const PORT = 4000;
59// Now that our HTTP server is fully set up, we can listen to it.
60httpServer.listen(PORT, () => {
61  console.log(`Server is now running on http://localhost:${PORT}/graphql`);
62});

⚠️ Running into an error? If you're using the graphql-ws library, your specified subscription protocol must be consistent across your backend, frontend, and every other tool you use (including Apollo Sandbox). For more information, see Switching from subscriptions-transport-ws in the Apollo Server 3 docs.

Resolving a subscription

Resolvers for Subscription fields differ from resolvers for fields of other types. Specifically, Subscription field resolvers are objects that define a subscribe function:

1const resolvers = {
2  Subscription: {
3    hello: {
4      // Example using an async generator
5      subscribe: async function* () {
6        for await (const word of ['Hello', 'Bonjour', 'Ciao']) {
7          yield { hello: word };
8        }
9      },
10    },
11    postCreated: {
12      // More on pubsub below
13      subscribe: () => pubsub.asyncIterator(['POST_CREATED']),
14    },
15  },
16  // ...other resolvers...
17};

The subscribe function must return an object of type AsyncIterator, a standard interface for iterating over asynchronous results. In the above postCreated.subscribe field, an AsyncIterator is generated by pubsub.asyncIterator (more on this below).

The PubSub class

The PubSub class is not recommended for production environments, because it's an in-memory event system that only supports a single server instance. After you get subscriptions working in development, we strongly recommend switching it out for a different subclass of the abstract PubSubEngine class. Recommended subclasses are listed in Production PubSub libraries.

You can use the publish-subscribe (pub/sub) model to track events that update active subscriptions. The graphql-subscriptions library provides the PubSub class as a basic in-memory event bus to help you get started:

To use the graphql-subscriptions package, first install it like so:

1npm install graphql-subscriptions

A PubSub instance enables your server code to both publish events to a particular label and listen for events associated with a particular label. We can create a PubSub instance like so:

1import { PubSub } from 'graphql-subscriptions';
2
3const pubsub = new PubSub();

Publishing an event

You can publish an event using the publish method of a PubSub instance:

1pubsub.publish('POST_CREATED', {
2  postCreated: {
3    author: 'Ali Baba',
4    comment: 'Open sesame',
5  },
6});

• The first parameter is the name of the event label you're publishing to, as a string.

  • You don't need to register a label name before publishing to it.

• The second parameter is the payload associated with the event.

  • The payload should include whatever data is necessary for your resolvers to populate the associated Subscription field and its subfields.

When working with GraphQL subscriptions, you publish an event whenever a subscription's return value should be updated. One common cause of such an update is a mutation, but any back-end logic might result in changes that should be published.

As an example, let's say our GraphQL API supports a createPost mutation:

1type Mutation {
2  createPost(author: String, comment: String): Post
3}

A basic resolver for createPost might look like this:

1const resolvers = {
2  Mutation: {
3    createPost(parent, args, { postController }) {
4      // Datastore logic lives in postController
5      return postController.createPost(args);
6    },
7  },
8  // ...other resolvers...
9};

Before we persist the new post's details in our datastore, we can publish an event that also includes those details:

1const resolvers = {
2  Mutation: {
3    createPost(parent, args, { postController }) {
4      pubsub.publish('POST_CREATED', { postCreated: args });
5      return postController.createPost(args);
6    },
7  },
8  // ...other resolvers...
9};

Next, we can listen for this event in our Subscription field's resolver.

Listening for events

An AsyncIterator object listens for events that are associated with a particular label (or set of labels) and adds them to a queue for processing.

You can create an AsyncIterator by calling the asyncIterator method of PubSub and passing in an array containing the names of the event labels that this AsyncIterator should listen for.

1pubsub.asyncIterator(['POST_CREATED']);

Every Subscription field resolver's subscribe function must return an AsyncIterator object.

This brings us back to the code sample at the top of Resolving a subscription:

1const resolvers = {
2  Subscription: {
3    postCreated: {
4      subscribe: () => pubsub.asyncIterator(['POST_CREATED']),
5    },
6  },
7  // ...other resolvers...
8};

With this subscribe function set, Apollo Server uses the payloads of POST_CREATED events to push updated values for the postCreated field.

Filtering events

Sometimes, a client should only receive updated subscription data if that data meets certain criteria. To support this, you can call the withFilter helper function in your Subscription field's resolver.

Example

Let's say our server provides a commentAdded subscription, which should notify clients whenever a comment is added to a specified code repository. A client can execute a subscription that looks like this:

1subscription ($repoName: String!) {
2  commentAdded(repoFullName: $repoName) {
3    id
4    content
5  }
6}

This presents a potential issue: our server probably publishes a COMMENT_ADDED event whenever a comment is added to any repository. This means that the commentAdded resolver executes for every new comment, regardless of which repository it's added to. As a result, subscribing clients might receive data they don't want (or shouldn't even have access to).

To fix this, we can use the withFilter helper function to control updates on a per-client basis.

Here's an example resolver for commentAdded that uses the withFilter function:

1import { withFilter } from 'graphql-subscriptions';
2
3const resolvers = {
4  Subscription: {
5    commentAdded: {
6      subscribe: withFilter(
7        () => pubsub.asyncIterator('COMMENT_ADDED'),
8        (payload, variables) => {
9          // Only push an update if the comment is on
10          // the correct repository for this operation
11          return (
12            payload.commentAdded.repository_name === variables.repoFullName
13          );
14        },
15      ),
16    },
17  },
18  // ...other resolvers...
19};

The withFilter function takes two parameters:

• The first parameter is exactly the function you would use for subscribe if you weren't applying a filter.

• The second parameter is a filter function that returns true if a subscription update should be sent to a particular client, and false otherwise (Promise<boolean> is also allowed). This function takes two parameters of its own:

  • payload is the payload of the event that was published.

  • variables is an object containing all arguments the client provided when initiating their subscription.

Use withFilter to make sure clients get exactly the subscription updates they want (and are allowed to receive).

Basic runnable example

An example server is available on GitHub and CodeSandbox:

The server exposes one subscription (numberIncremented) that returns an integer that's incremented on the server every second. Here's an example subscription that you can run against your server:

1subscription IncrementingNumber {
2  numberIncremented
3}

If you don't see the result of your subscription operation you might need to adjust your Sandbox settings to use the graphql-ws protocol.

After you start up the server in CodeSandbox, follow the instructions in the browser to test running the numberIncremented subscription in Apollo Sandbox. You should see the subscription's value update every second.

Operation context

When initializing context for a query or mutation, you usually extract HTTP headers and other request metadata from the req object provided to the context function.

For subscriptions, you can extract information from a client's request by adding options to the first argument passed to the useServer function.

For example, you can provide a context property to add values to your GraphQL operation contextValue:

1// ...
2useServer(
3  {
4    // Our GraphQL schema.
5    schema,
6    // Adding a context property lets you add data to your GraphQL operation contextValue
7    context: async (ctx, msg, args) => {
8      // You can define your own function for setting a dynamic context
9      // or provide a static value
10      return getDynamicContext(ctx, msg, args);
11    },
12  },
13  wsServer,
14);

Notice that the first parameter passed to the context function above is ctx. The ctx object represents the context of your subscription server, not the GraphQL operation contextValue that's passed to your resolvers.

You can access the parameters of a client's subscription request to your WebSocket server via the ctx.connectionParams property.

Below is an example of the common use case of extracting an authentication token from a client subscription request and using it to find the current user:

1const getDynamicContext = async (ctx, msg, args) => {
2  // ctx is the graphql-ws Context where connectionParams live
3  if (ctx.connectionParams.authentication) {
4    const currentUser = await findUser(ctx.connectionParams.authentication);
5    return { currentUser };
6  }
7  // Otherwise let our resolvers know we don't have a current user
8  return { currentUser: null };
9};
10
11useServer(
12  {
13    schema,
14    context: async (ctx, msg, args) => {
15      // Returning an object will add that information to
16      // contextValue, which all of our resolvers have access to.
17      return getDynamicContext(ctx, msg, args);
18    },
19  },
20  wsServer,
21);

Putting it all together, the useServer.context function returns an object, contextValue, which is available to your resolvers.

Note that the context option is called once per subscription request, not once per event emission. This means that in the above example, every time a client sends a subscription operation, we check their authentication token.

onConnect and onDisconnect

You can configure the subscription server's behavior whenever a client connects (onConnect) or disconnects (onDisconnect).

Defining an onConnect function enables you to reject a particular incoming connection by returning false or by throwing an exception. This can be helpful if you want to check authentication when a client first connects to your subscription server.

You provide these functions as options to the first argument of useServer, like so:

1useServer(
2  {
3    schema,
4    // As before, ctx is the graphql-ws Context where connectionParams live.
5    onConnect: async (ctx) => {
6      // Check authentication every time a client connects.
7      if (tokenIsNotValid(ctx.connectionParams)) {
8        // You can return false to close the connection  or throw an explicit error
9        throw new Error('Auth token missing!');
10      }
11    },
12    onDisconnect(ctx, code, reason) {
13      console.log('Disconnected!');
14    },
15  },
16  wsServer,
17);

For more examples of using onConnect and onDisconnect, see the graphql-ws recipes documentation.

Example: Authentication with Apollo Client

If you plan to use subscriptions with Apollo Client, ensure both your client and server subscription protocols are consistent for the subscription library you're using (this example uses the graphql-ws library).

In Apollo Client, the GraphQLWsLink constructor supports adding information to connectionParams (example). Those connectionParams are sent to your server when it connects, allowing you to extract information from the client request by accessing Context.connectionParams.

Let's suppose we create our subscription client like so:

1import { GraphQLWsLink } from '@apollo/client/link/subscriptions';
2import { createClient } from 'graphql-ws';
3
4const wsLink = new GraphQLWsLink(
5  createClient({
6    url: 'ws://localhost:4000/subscriptions',
7    connectionParams: {
8      authentication: user.authToken,
9    },
10  }),
11);

The connectionParams argument (which contains the information provided by the client) is passed to your server, enabling you to validate the user's credentials.

From there you can use the useServer.context property to authenticate the user, returning an object that's passed into your resolvers as the context argument during execution.

For our example, we can use the connectionParams.authentication value provided by the client to look up the related user before passing that user along to our resolvers:

1const findUser = async (authToken) => {
2  // Find a user by their auth token
3};
4
5const getDynamicContext = async (ctx, msg, args) => {
6  if (ctx.connectionParams.authentication) {
7    const currentUser = await findUser(ctx.connectionParams.authentication);
8    return { currentUser };
9  }
10  // Let the resolvers know we don't have a current user so they can
11  // throw the appropriate error
12  return { currentUser: null };
13};
14
15// ...
16useServer(
17  {
18    // Our GraphQL schema.
19    schema,
20    context: async (ctx, msg, args) => {
21      // This will be run every time the client sends a subscription request
22      return getDynamicContext(ctx, msg, args);
23    },
24  },
25  wsServer,
26);

To sum up, the example above looks up a user based on the authentication token sent with each subscription request before returning the user object to be used by our resolvers. If no user exists or the lookup otherwise fails, our resolvers can throw an error and the corresponding subscription operation is not executed.

Production PubSub libraries

As mentioned above, the PubSub class is not recommended for production environments, because its event-publishing system is in-memory. This means that events published by one instance of your GraphQL server are not received by subscriptions that are handled by other instances.

Instead, you should use a subclass of the PubSubEngine abstract class that you can back with an external datastore such as Redis or Kafka.

The following are community-created PubSub libraries for popular event-publishing systems:

• Solace

• Redis

• Google PubSub

• MQTT enabled broker

• RabbitMQ

• AMQP (RabbitMQ)

• Kafka

• Postgres

• Postgres With Typescript

• Google Cloud Firestore

• Ably Realtime

• Google Firebase Realtime Database

• Azure SignalR Service

• Azure ServiceBus

• MongoDB

If none of these libraries fits your use case, you can also create your own PubSubEngine subclass. If you create a new open-source library, click Edit on GitHub to let us know about it!