Implementing a Gateway with Apollo Server

After you set up at least one federation-ready subgraph, you can configure a graph router (also known as a gateway) to sit in front of your subgraphs.

📣 In the majority of cases, we recommend using the GraphOS Router as your graph router. It's faster to configure, it's more performant (especially with high request loads), and it rarely requires writing custom code.

In certain cases, you might need to use Apollo Server as your graph router if your subgraphs use a custom authentication method that is currently difficult to configure with the GraphOS Router.

Regardless of which graph router library you start with, you can swap to the other without making any changes to other parts of your supergraph.

Node.js gateway setup

This section walks through setting up a basic graph router using Apollo Server and the @apollo/gateway library. It currently requires Node.js version 14 or 16.

Create a new Node.js project with npm init, then install the necessary packages:

1npm install @apollo/gateway @apollo/server graphql

The @apollo/gateway package includes the ApolloGateway class. To configure Apollo Server to act as a graph router, you pass an instance of ApolloGateway to the ApolloServer constructor, like so:

1import { ApolloServer } from '@apollo/server';
2import { startStandaloneServer } from '@apollo/server/standalone';
3import { ApolloGateway } from '@apollo/gateway';
4import { readFileSync } from 'fs';
5
6const supergraphSdl = readFileSync('./supergraph.graphql').toString();
7
8// Initialize an ApolloGateway instance and pass it
9// the supergraph schema as a string
10const gateway = new ApolloGateway({
11  supergraphSdl,
12});
13
14// Pass the ApolloGateway to the ApolloServer constructor
15const server = new ApolloServer({
16  gateway,
17});
18
19// Note the top-level `await`!
20const { url } = await startStandaloneServer(server);
21console.log(`🚀  Server ready at ${url}`);

Composing the supergraph schema

In the above example, we provide the supergraphSdl option to the ApolloGateway constructor. This is the string representation of our supergraph schema, which is composed from all of our subgraph schemas.

To learn how to compose your supergraph schema, see Supported methods.

In production, we strongly recommend running the gateway in a managed mode with Apollo Studio, which enables your gateway to update its configuration without a restart. For details, see Setting up managed federation.

On startup, the gateway processes your supergraphSdl, which includes routing information for your subgraphs. It then begins accepting incoming requests and creates query plans for them that execute across one or more subgraphs.

Updating the supergraph schema

In the above example, we provide a static supergraph schema to the gateway. This approach requires the gateway to restart in order to update the supergraph schema. This is undesirable for many applications, so we also provide the ability to update the supergraph schema dynamically.

1import { ApolloServer } from '@apollo/server';
2import { startStandaloneServer } from '@apollo/server/standalone';
3import { ApolloGateway } from '@apollo/gateway';
4import { readFile } from 'fs/promises';
5
6let supergraphUpdate;
7const gateway = new ApolloGateway({
8  async supergraphSdl({ update }) {
9    // `update` is a function that we'll save for later use
10    supergraphUpdate = update;
11    return {
12      supergraphSdl: await readFile('./supergraph.graphql', 'utf-8'),
13    };
14  },
15});
16
17// Pass the ApolloGateway to the ApolloServer constructor
18const server = new ApolloServer({
19  gateway,
20});
21
22const { url } = await startStandaloneServer(server);
23console.log(`🚀  Server ready at ${url}`);

There are a few things happening here. Let's take a look at each of them individually.

Note that supergraphSdl is now an async function. This function is called exactly once, when ApolloServer initializes the gateway. It has the following responsibilities:

• It receives the update function, which we use to update the supergraph schema.

• It returns the initial supergraph schema, which the gateway uses at startup.

With the update function, we can now programmatically update the supergraph schema. Polling, webhooks, and file watchers are all examples of ways we can go about doing this.

The code below demonstrates a more complete example using a file watcher. In this example, assume that we're updating the supergraphSdl.graphql file with the Rover CLI.

1import { ApolloServer } from '@apollo/server';
2import { startStandaloneServer } from '@apollo/server/standalone';
3import { ApolloGateway } from '@apollo/gateway';
4import { watch } from 'fs';
5import { readFile } from 'fs/promises';
6
7const server = new ApolloServer({
8  gateway: new ApolloGateway({
9    async supergraphSdl({ update, healthCheck }) {
10      // create a file watcher
11      const watcher = watch('./supergraph.graphql');
12      // subscribe to file changes
13      watcher.on('change', async () => {
14        // update the supergraph schema
15        try {
16          const updatedSupergraph = await readFile(
17            './supergraph.graphql',
18            'utf-8',
19          );
20          // optional health check update to ensure our services are responsive
21          await healthCheck(updatedSupergraph);
22          // update the supergraph schema
23          update(updatedSupergraph);
24        } catch (e) {
25          // handle errors that occur during health check or while updating the supergraph schema
26          console.error(e);
27        }
28      });
29
30      return {
31        supergraphSdl: await readFile('./supergraph.graphql', 'utf-8'),
32        // cleanup is called when the gateway is stopped
33        async cleanup() {
34          watcher.close();
35        },
36      };
37    },
38  }),
39});
40
41const { url } = await startStandaloneServer(server);
42console.log(`🚀  Server ready at ${url}`);

This example is a bit more complete. Let's take a look at what we've added.

In the supergraphSdl callback, we also receive a healthCheck function. This enables us to run a health check against each of the services in our future supergraph schema. This is useful for ensuring that our services are responsive and that we don't perform an update when it's unsafe.

We've also wrapped our call to update and healthCheck in a try block. If an error occurs during either of these, we want to handle this gracefully. In this example, we continue running the existing supergraph schema and log an error.

Finally, we return a cleanup function. This is a callback that's called when the gateway is stopped. This enables us to cleanly shut down any ongoing processes (such as file watching or polling) when the gateway is shut down via a call to ApolloServer.stop. The gateway expects cleanup to return a Promise and awaits it before shutting down.

Advanced usage

In a more complex application, you might want to create a class that handles the update and healthCheck functions, along with any additional state. In this case, you can instead provide an object (or class) with an initialize function. This function is called just like the supergraphSdl function discussed above. For an example of this, see the IntrospectAndCompose source code.

Composing subgraphs with IntrospectAndCompose

⚠️ We strongly recommend against using IntrospectAndCompose in production. For details, see Limitations of IntrospectAndCompose.

Instead of providing a composed supergraph schema to the gateway, you can instruct the gateway to fetch all of your subgraph schemas and perform composition itself. To do so, provide an instance of the IntrospectAndCompose class with a subgraphs array, like so:

1const { ApolloGateway, IntrospectAndCompose } = require('@apollo/gateway');
2
3const gateway = new ApolloGateway({
4  supergraphSdl: new IntrospectAndCompose({
5    subgraphs: [
6      { name: 'accounts', url: 'http://localhost:4001' },
7      { name: 'products', url: 'http://localhost:4002' },
8      // ...additional subgraphs...
9    ],
10  }),
11});

Each item in the subgraphs array is an object that specifies the name and url of one of your subgraphs. You can specify any string value for name, which is used primarily for query planner output, error messages, and logging.

On startup, the gateway fetches each subgraph's schema from its url and composes those schemas into a supergraph schema. It then begins accepting incoming requests and creates query plans for them that execute across one or more subgraphs.

Additional configuration options can be found in the IntrospectAndCompose API documentation.

However, IntrospectAndCompose has important limitations.

Limitations of IntrospectAndCompose

The IntrospectAndCompose option can sometimes be helpful for local development, but it's strongly discouraged for any other environment. Here are some reasons why:

• Composition might fail. With IntrospectAndCompose, your gateway performs composition dynamically on startup, which requires network communication with each subgraph. If composition fails, your gateway throws errors and experiences unplanned downtime.

  • With the static or dynamic supergraphSdl configuration, you instead provide a supergraph schema that has already been composed successfully. This prevents composition errors and enables faster startup.

• Gateway instances might differ. If you deploy multiple instances of your gateway while deploying updates to your subgraphs, your gateway instances might fetch different schemas from the same subgraph. This can result in sporadic composition failures or inconsistent supergraph schemas between instances.

  • When you deploy multiple instances with supergraphSdl, you provide the exact same static artifact to each instance, enabling more predictable behavior.

Updating the gateway

Before updating your gateway's version, check the changelog for potential breaking changes.

We strongly recommend updating your gateway in local and test environments before deploying updates to staging or production.

You can confirm the currently installed version of the @apollo/gateway library with the npm list command:

1npm list @apollo/gateway

To update the library, use the npm update command:

1npm update @apollo/gateway

This updates the library to the most recent version allowed by your package.json file. Learn more about dependency constraints.

To update to a particular version (including a version that exceeds your dependency constraints), use npm install instead:

1npm install @apollo/[email protected]

Customizing requests and responses

The gateway can modify the details of an incoming request before executing it across your subgraphs. For example, your subgraphs might all use the same authorization token to associate an incoming request with a particular user. The gateway can add that token to each operation it sends to your subgraphs.

Similarly, the gateway can modify the details of its response to a client, based on the result returned by each subgraph.

Customizing requests

In the following example, each incoming request to the gateway includes an Authorization header. The gateway sets the shared contextValue for an operation by pulling the value of that header and using it to fetch the associated user's ID.

After adding the userId to the shared contextValue object, the gateway can then add that value to a header that it includes in its requests to each subgraph.

1import { ApolloServer } from '@apollo/server';
2import { startStandaloneServer } from '@apollo/server/standalone';
3import { ApolloGateway, RemoteGraphQLDataSource } from '@apollo/gateway';
4import { readFileSync } from 'fs';
5
6class AuthenticatedDataSource extends RemoteGraphQLDataSource {
7  willSendRequest({ request, context }) {
8    // Pass the user's id from the context to each subgraph
9    // as a header called `user-id`
10    request.http.headers.set('user-id', context.userId);
11  }
12}
13
14const supergraphSdl = readFileSync('./supergraph.graphql').toString();
15const gateway = new ApolloGateway({
16  supergraphSdl,
17  buildService({ name, url }) {
18    return new AuthenticatedDataSource({ url });
19  },
20});
21
22const server = new ApolloServer({
23  gateway,
24});
25
26const { url } = await startStandaloneServer(server, {
27  context: async ({ req }) => {
28    // Get the user token from the headers
29    const token = req.headers.authorization || '';
30    // Try to retrieve a user with the token
31    const userId = getUserId(token);
32    // Add the user ID to the contextValue
33    return { userId };
34  },
35});
36
37console.log(`🚀  Server ready at ${url}`);

The fields of the object passed to your context function might differ if you're using a different Apollo Server integration.

The buildService function enables us to customize the requests that are sent to our subgraphs. In this example, we return a custom RemoteGraphQLDataSource. This datasource enables us to modify the outgoing request with information from Apollo Server's contextValue before it's sent. Here, we add the user-id header to pass an authenticated user ID to downstream services.

Customizing responses

Let's say that whenever a subgraph returns an operation result to the gateway, it includes a Server-Id header in the response. The value of the header uniquely identifies the subgraph in our graph.

When the gateway then responds to a client, we want its Server-Id header to include the identifier for every subgraph that contributed to the response. In this case, we can tell the gateway to aggregate the various server IDs into a single, comma-separated list.

The flow for processing a single operation from a client application then looks like this:

To implement this flow, we can use the didReceiveResponse callback of the RemoteGraphQLDataSource class to inspect each subgraph's result as it comes in. We can add the Server-Id to the shared context in this callback, then pull the full list from the context when sending the final response to the client.

1import { ApolloServer } from '@apollo/server';
2import { startStandaloneServer } from '@apollo/server/standalone';
3import { ApolloGateway, RemoteGraphQLDataSource } from '@apollo/gateway';
4import { readFileSync } from 'fs';
5
6const supergraphSdl = readFileSync('./supergraph.graphql').toString();
7
8class DataSourceWithServerId extends RemoteGraphQLDataSource {
9  async didReceiveResponse({ response, request, context }) {
10    // Parse the Server-Id header and add it to the array on context
11    const serverId = response.http.headers.get('Server-Id');
12    if (serverId) {
13      context.serverIds.push(serverId);
14    }
15
16    // Return the response, even when unchanged.
17    return response;
18  }
19}
20
21const gateway = new ApolloGateway({
22  supergraphSdl,
23  buildService({ url }) {
24    return new DataSourceWithServerId({ url });
25  },
26});
27
28const server = new ApolloServer({
29  gateway,
30  plugins: [
31    {
32      requestDidStart() {
33        return {
34          async willSendResponse({ contextValue, response }) {
35            // Append our final result to the outgoing response headers
36            response.http.headers.set(
37              'Server-Id',
38              contextValue.serverIds.join(','),
39            );
40          },
41        };
42      },
43    },
44  ],
45});
46
47const { url } = await startStandaloneServer(server, {
48  context() {
49    return { serverIds: [] };
50  },
51});
52
53console.log(`🚀  Server ready at ${url}`);

In this example, multiple calls to didReceiveResponse are pushing a value onto the shared contextValue.serverIds array. The order of these calls cannot be guaranteed. If you write logic that modifies the shared contextValue object, make sure that modifications are not destructive, and that the order of modifications doesn't matter.

To learn more about buildService and RemoteGraphQLDataSource, see the API docs.

Custom directive support

The @apollo/gateway library supports the use of custom directives in your subgraph schemas. This support differs depending on whether a given directive is a type system directive or an executable directive.

Type system directives

Type system directives are directives that are applied to one of these locations. These directives are not used within operations, but rather are applied to locations within the schema itself.

The @deprecated directive below is an example of a type system directive:

1directive @deprecated(
2  reason: String = "No longer supported"
3) on FIELD_DEFINITION | ENUM_VALUE
4
5type ExampleType {
6  newField: String
7  oldField: String @deprecated(reason: "Use `newField`.")
8}

The gateway strips all definitions and uses of type system directives from your graph's API schema. This has no effect on your subgraph schemas, which retain this information.

Effectively, the gateway supports type system directives by ignoring them, making them the responsibility of the subgraphs that define them.

Executable directives

Executable directives are directives that are applied to one of these locations. These directives are defined in your schema, but they're used in operations that are sent by clients.

Here's an example of an executable directive definition:

1# Uppercase this field's value (assuming it's a string)
2directive @uppercase on FIELD

And here's an example of a query that uses that directive:

1query GetUppercaseUsernames {
2  users {
3    name @uppercase
4  }
5}

It's strongly recommended that all of your subgraphs use the exact same logic for a given executable directive. Otherwise, operations might produce inconsistent or confusing results for clients.