GraphQL

Global Object Identification

Consistent object access enables simple caching and object lookups

To provide options for GraphQL clients to elegantly handle caching and data refetching, GraphQL servers need to expose object identifiers in a standardized way.

For this to work, a client will need to query via a standard mechanism to request an object by ID. Then, in the response, the schema will need to provide a standard way of providing these IDs.

Because little is known about the object other than its ID, we call these objects "nodes." Here is an example query for a node:

{
  node(id: "4") {
    id
    ... on User {
      name
    }
  }
  • The GraphQL schema is formatted to allow fetching any object via the node field on the root query object. This returns objects which conform to a "Node" interface.
  • The id field can be extracted out of the response safely, and can be stored for re-use via caching and refetching.
  • Clients can use interface fragments to extract additional information specific to the type which conform to the node interface. In this case a "User".

The Node interface looks like:

# An object with a Globally Unique ID
interface Node {
  # The ID of the object.
  id: ID!
}

With a User conforming via:

type User implements Node {
  id: ID!
  # Full name
  name: String!
}

Specification #

Everything below describes with more formal requirements a specification around object identification in order to conform to ensure consistency across server implementations. These specifications are based on how a server can be compliant with the Relay API client, but can be useful for any client.

Reserved Types #

A GraphQL server compatible with this spec must reserve certain types and type names to support the consistent object identification model. In particular, this spec creates guidelines for the following types:

  • An interface named Node.
  • The node field on the root query type.

Node Interface #

The server must provide an interface called Node. That interface must include exactly one field, called id that returns a non-null ID.

This id should be a globally unique identifier for this object, and given just this id, the server should be able to refetch the object.

Introspection #

A server that correctly implements the above interface will accept the following introspection query, and return the provided response:

{
  __type(name: "Node") {
    name
    kind
    fields {
      name
      type {
        kind
        ofType {
          name
          kind
        }
      }
    }
  }
}

yields

{
  "__type": {
    "name": "Node",
    "kind": "INTERFACE",
    "fields": [
      {
        "name": "id",
        "type": {
          "kind": "NON_NULL",
          "ofType": {
            "name": "ID",
            "kind": "SCALAR"
          }
        }
      }
    ]
  }
}

Node root field #

The server must provide a root field called node that returns the Node interface. This root field must take exactly one argument, a non-null ID named id.

If a query returns an object that implements Node, then this root field should refetch the identical object when value returned by the server in the Node's id field is passed as the id parameter to the node root field.

The server must make a best effort to fetch this data, but it may not always be possible; for example, the server may return a User with a valid id, but when the request is made to refetch that user with the node root field, the user's database may be unavailable, or the user may have deleted their account. In this case, the result of querying this field should be null.

Introspection #

A server that correctly implements the above requirement will accept the following introspection query, and return a response that contains the provided response.

{
  __schema {
    queryType {
      fields {
        name
        type {
          name
          kind
        }
        args {
          name
          type {
            kind
            ofType {
              name
              kind
            }
          }
        }
      }
    }
  }
}

yields

{
  "__schema": {
    "queryType": {
      "fields": [
        // This array may have other entries
        {
          "name": "node",
          "type": {
            "name": "Node",
            "kind": "INTERFACE"
          },
          "args": [
            {
              "name": "id",
              "type": {
                "kind": "NON_NULL",
                "ofType": {
                  "name": "ID",
                  "kind": "SCALAR"
                }
              }
            }
          ]
        }
      ]
    }
  }
}

Field stability #

If two objects appear in a query, both implementing Node with identical IDs, then the two objects must be equal.

For the purposes of this definition, object equality is defined as follows:

  • If a field is queried on both objects, the result of querying that field on the first object must be equal to the result of querying that field on the second object.
    • If the field returns a scalar, equality is defined as is appropriate for that scalar.
    • If the field returns an enum, equality is defined as both fields returning the same enum value.
    • If the field returns an object, equality is defined recursively as per the above.

For example:

{
  fourNode: node(id: "4") {
    id
    ... on User {
      name
      userWithIdOneGreater {
        id
        name
      }
    }
  }
  fiveNode: node(id: "5") {
    id
    ... on User {
      name
      userWithIdOneLess {
        id
        name
      }
    }
  }
}

might return:

{
  "fourNode": {
    "id": "4",
    "name": "Mark Zuckerberg",
    "userWithIdOneGreater": {
      "id": "5",
      "name": "Chris Hughes"
    }
  },
  "fiveNode": {
    "id": "5",
    "name": "Chris Hughes",
    "userWithIdOneLess": {
      "id": "4",
      "name": "Mark Zuckerberg",
    }
  }
}

Because fourNode.id and fiveNode.userWithIdOneLess.id are the same, we are guaranteed by the conditions above that fourNode.name must be the same as fiveNode.userWithIdOneLess.name, and indeed it is.

Plural identifying root fields #

Imagine a root field named username, that takes a user's username and returns the corresponding user:

{
  username(username: "zuck") {
    id
  }
}

might return:

{
  "username": {
    "id": "4",
  }
}

Clearly, we can link up the object in the response, the user with ID 4, with the request, identifying the object with username "zuck". Now imagine a root field named usernames, that takes a list of usernames and returns a list of objects:

{
  usernames(usernames: ["zuck", "moskov"]) {
    id
  }
}

might return:

{
  "usernames": [
    {
      "id": "4",
    },
    {
      "id": "6"
    }
  ]
}

For clients to be able to link the usernames to the responses, it needs to know that the array in the response will be the same size as the array passed as an argument, and that the order in the response will match the order in the argument. We call these plural identifying root fields, and their requirements are described below.

Fields #

A server compliant with this spec may expose root fields that accept a list of input arguments, and returns a list of responses. For spec-compliant clients to use these fields, these fields must be plural identifying root fields, and obey the following requirements.

NOTE Spec-compliant servers may expose root fields that are not plural identifying root fields; the spec-compliant client will just be unable to use those fields as root fields in its queries.

Plural identifying root fields must have a single argument. The type of that argument must be a non-null list of non-nulls. In our usernames example, the field would take a single argument named usernames, whose type (using our type system shorthand) would be [String!]!.

The return type of a plural identifying root field must be a list, or a non-null wrapper around a list. The list must wrap the Node interface, an object that implements the Node interface, or a non-null wrapper around those types.

Whenever the plural identifying root field is used, the length of the list in the response must be the same as the length of the list in the arguments. Each item in the response must correspond to its item in the input; more formally, if passing the root field an input list Lin resulted in output value Lout, then for an arbitrary permutation P, passing the root field P(Lin) must result in output value P(Lout).

Because of this, servers are advised to not have the response type wrap a non-null wrapper, because if it is unable to fetch the object for a given entry in the input, it still must provide a value in the output for that input entry; null is a useful value for doing so.

Continue Reading →Caching