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 } }}
node
field on the root query object. This
returns objects which conform to a "Node" interface.id
field can be extracted out of the response safely, and can be stored for re-use via caching and refetching.The Node interface looks like:
# An object with a Globally Unique IDinterface Node { # The ID of the object. id: ID!}
With a User conforming via:
type User implements Node { id: ID! # Full name name: String!}
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.
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:
Node
.node
field on the root query type.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.
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" } } } ] }}
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
.
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" } } } ] } ] } }}
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:
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.
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.
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.