/
Launch Apollo Studio

The gateway


After you set up at least one federation-ready subgraph, you can configure a gateway to sit in front of your subgraphs. The gateway executes incoming operations across those subgraphs.

The @apollo/gateway package extends Apollo Server's functionality, enabling it to act as a gateway for an Apollo Federation architecture.

We recommend against running your gateway in a serverless environment (such as AWS Lambda), because schema composition increases Apollo Server's startup time. If you do run your gateway in a serverless environment, set your function's timeout to at least 10 seconds to prevent unexpected errors.

Setup

First, let's install the necessary packages:

npm install @apollo/gateway apollo-server graphql

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

const { ApolloServer } = require('apollo-server');
const { ApolloGateway } = require('@apollo/gateway');

// Initialize an ApolloGateway instance and pass it an array of
// your subgraph names and URLs
const gateway = new ApolloGateway({
  serviceList: [
    { name: 'accounts', url: 'http://localhost:4001' },
    // Define additional services here
  ],
});

// Pass the ApolloGateway to the ApolloServer constructor
const server = new ApolloServer({
  gateway,

  // Disable subscriptions (not currently supported with ApolloGateway)
  subscriptions: false,
});

server.listen().then(({ url }) => {
  console.log(`🚀 Server ready at ${url}`);
});

In the above example, we provide the serviceList configuration option to the ApolloGateway constructor. This array specifies a name and url for each of our subgraphs. You can specify any string value for name, which is used primarily for query planner output, error messages, and logging.

In production, we recommend running the gateway in a managed mode with Apollo Studio, which relies on static files rather than introspection. For details, see Setting up managed federation.

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

If there are any composition errors, the new ApolloServer call throws an exception with a list of validation errors.

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 context 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 context object, the gateway can then add that value to a header that it includes in its requests to each subgraph.

The fields of the object passed to your context function differ if you're using middleware besides Express. See the API reference for details.

The buildService function enables us to customize the requests that are sent to our subgraphs. In this example, we return a custom RemoteGraphQLDataSource. The datasource allows us to modify the outgoing request with information from the Apollo Server context 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:

Client appGatewaySubgraphsSends GraphQL operationGenerates query plan for operationSends the operation to the applicable subgraphResponds with result and Server-Id headerAdds the returned Server-Id to the shared contextloop[For each operation in the query plan]Adds all Server-Ids in the shared context to the response headerSends operation responseClient appGatewaySubgraphs

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.

In this example, multiple calls to didReceiveResponse are pushing a value onto the shared context.serverIds array. The order of these calls cannot be guaranteed. If you write logic that modifies the shared context 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:

directive @deprecated(
  reason: String = "No longer supported"
) on FIELD_DEFINITION | ENUM_VALUE

type ExampleType {
  newField: String
  oldField: String @deprecated(reason: "Use `newField`.")
}

At composition time, ApolloGateway strips all definitions and uses of type system directives from your composed 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.

Although the @apollo/gateway library supports executable directives, Apollo Server itself does not. This guidance is provided primarily for architectures that use the @apollo/gateway library in combination with subgraphs that do not use Apollo Server.

Here's an example of an executable directive definition:

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

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

query GetUppercaseUsernames {
  users {
    name @uppercase
  }
}

At composition time, ApolloGateway makes sure that all of your subgraphs define the exact same set of executable directives. If any service is missing a definition, or if definitions differ in their locations, arguments, or argument types, a composition error occurs.

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

Edit on GitHub