Andreas Garnæs

Type-Safe GraphQL with OCaml (part 1)

In July 2016, I spent some time writing a GraphQL endpoint for an existing application in Go. I chose to implement the endpoint in Go based on prior experience, good concurrency support and the desire to have static type system. The GraphQL library for Go, graphql-go, worked pretty well, and I was able to build the desired functionality. Still, I was left frustrated by the lack of type safety offered in Go. Everywhere the library neccessitated the use of interface{} types, and the application was littered with type casts. Code like the following was the rule rather than the exception:

func(p graphql.ResolveParams) (interface{}, error) {
  ctx := p.Context.(AppContext)           // Type cast 1
  account := p.Source.(Account)           // Type cast 2
  userId := p.Args["user_id"].(int)       // Type cast 3

  return ctx.LoadUser(account.Id, userId) // Any return value accepted
}

Ugh, three type casts to implement a very simple function. Secondly, the return type is interface{}, so there’s no type safety for the return value. Type casts and lack of type safety collectively mean a much higher risk of runtime errors.

Initially, I thought the lack of generics in Go was the root of the issue. Out of curiosity, I looked into graphql-java, and noticed that it offered equally bad type safety guarantees. I went on to look at Flow to see if it was possible to get typing for graphql-js, but unfortunately it did not (and still does not) support GraphQL (Github issue). This was the starting point of a journey to try to implement a GraphQL library with better type safety guarantees than were offered at the time.

My intent is to write a number of blog posts, which describe that journey. Incidentally, it also describes the implementation of ocaml-graphql-server, which has become the product of the journey. This blog post, part 1, provides a foundation for understanding GraphQL from a server perspective, and describes a type-safe implementation of a subset of GraphQL in OCaml.

GraphQL Primer

This section provides a primer to a simplified version of GraphQL, which will provide a foundation for later describing how to implement a GraphQL server library.

GraphQL requires you to model your data as a directed graph. As a very simple example, we might have a user with two properties, id and name:

To capture the structure of this graph, we need to describe nodes, edges and the relationship between the two more accurately:

Nodes are collectively called types in GraphQL, i.e. scalars and objects are GraphQL types, and you will also find the terms “scalar type” and “object type” being used.

Example
Int is a scalar type with source type int. User is an object type with source type user. A field (edge) from User to Int must have a resolve function of type user -> int.

One node in the graph is designated as the query root (named query in the above example). The root node must be an object with “nothing” as its source type (i.e. unit in OCaml, void in C/Java, etc). A query is then a traversal in the graph starting at the root. The result of a query is a JSON value, which can be computed by running the following recursive procedure:

From an inductive perspective, scalars are the base case, while objects require recursive application.

Example
Executing a query on the above graph could yield the following JSON value:

{
  "user": {
    "id": 1,
    "name": "Alice"
  }
}

A final and important aspect of GraphQL is introspection. A client can issue a query to a GraphQL API asking for its structure and receive a JSON result describing the graph, i.e. the nodes and edges. We will omit the exact details of this format, but provide the following example.

Example
Introspecting the above example graph could yield a JSON value like the following:

{
  "queryType": "Query",
  "types": [
    {
      "name": "Query",
      "kind": "OBJECT",
      "fields": [
        {
          "name": "user",
          "type": "User"
        }
      ]
    },
    {
      "name": "User",
      "kind": "OBJECT",
      "fields": [
        {
          "name": "id",
          "type": "Int"
        },
        {
          "name": "name",
          "type": "String"
        }
      ]
    },
    {
      "name": "Int",
      "kind": "SCALAR"
    },
    {
      "name": "String",
      "kind": "SCALAR"
    }
  ]
}

Implementing the Core of GraphQL

Broadly speaking, the goal of a GraphQL server library is to:

  1. Allow the user to construct a schema (graph) in terms of objects, scalars and fields with application-specific serialize and resolve functions.
  2. Allow executing a query against a schema.
  3. Allow introspection of a schema.

The challenge is to define OCaml types for GraphQL types and fields, which capture the above requirements. Furthermore, the definition needs to support introspection, so we cannot just have a graph of closures, which hide the structure of the graph itself. Finally, the following invariant for fields should be upheld to guarantee that the schema is well-formed:

Field Invariant
Given a field with a resolve function of type x -> y, the source type of the tail must be x and the source type of the head must be y.

If we can capture this invariant in the type system, then only well-formed GraphQL schemas will compile and we can avoid exceptions when executing a query.

As a running example, we’ll try to expose a simple user type as a GraphQL object:

type user = {
  id   : int;
  name : string;
}

We’ll be using the following simple definition of a JSON type:

type json =
  | Null
  | Int    of int
  | Float  of float
  | String of string
  | Bool   of bool
  | Array  of json list
  | Object of (string * json) list

Given this JSON type, we could implement a simple conversion from user to json as follows:

(* user_to_json : user -> json *)
let user_to_json user =
  Object [
    ("id",   Int user.id);
    ("name", String user.name);
  ]

On to implementing our GraphQL library! A GraphQL type is parameterized over a source type, so we will have a type like 'src typ (type is a reserved keyword, so we use typ instead). First we will try to tackle scalars and then add objects. Scalars have a name and a serialize function to convert their source type into JSON:

module Graphql = struct
  type 'src typ =
    | Scalar of {
        name      : string;
        serialize : 'src -> json;
      }
end

GraphQL defines a number of built-in scalars, e.g. string and int, which we can define as follows using the above definition:

module Graphql = struct
  type 'src typ = (* ... *)

  (* int Graphql.typ *)
  let int = Scalar {
    name      = "Int";
    serialize = fun i -> Int i;
  }

  (* string Graphql.typ *)
  let string = Scalar {
    name      = "String";
    serialize = fun s -> String s;
  }
end

Before expanding our definition of Graphql.typ to include objects, let’s try to define a field. A field has a name, an output type, and a resolve function:

module Graphql = struct
  type 'src typ = (* ... *)

  and ('src, 'out) field = {
    name        : string;
    output_type : 'out typ;
    resolve     : 'src -> 'out;
  }
end

The type ('src, 'out) field can be interpreted as a field going from a node with source type 'src to a node with source type 'out. Note that the return value of resolve agrees with the source type of the field output_type (our field invariant).

We can then try to define the two fields of our user-type:

(* (user, int) Graphql.field *)
let id_field = Graphql.Field {
  name       = "id";
  output_typ = Graphql.int;
  resolve    = fun user -> user.id
}

(* (user, string) Graphql.field *)
let name_field = Graphql.Field {
  name       = "name";
  output_typ = Graphql.string;
  resolve    = fun user -> user.name
}

Though these definitions might seem fine, the fact that they have a different types might cause concern for the observant reader. In particular, this means we cannot put them in the same list:

(* TYPE ERROR! *)
let user_fields = [id_field; name_field] 

The implication is that we cannot construct a list of fields for an object with the above field types. The insight to overcome this issue is that while we need 'out typ and 'src -> 'out to agree, we otherwise do not care about 'out. In particular, 'out does not need to be exposed in the type of field. We can achieve this using a GADT:

module Graphql = struct
  type 'src typ = (* ... *)

  and 'src field = Field : {
    name        : string;
    output_type : 'out typ;
    resolve     : 'src -> 'out;
  } -> 'src field
end

By “forgetting” 'out, the type of id_field and name_field are now both of type user Graphql.field. This means we can create a list containing both user fields:

(* user Graphql.field list *)
let user_fields = [id_field; name_field]

With the above in place, we have made it really easy to extend our Graphql.typ to include object types. An object is simply a name and a list of fields that agree with the source type of the object:

module Graphql = struct
  type 'src typ =
    | Scalar of {
        name      : string;
        serialize : 'src -> json;
      }
    | Object of {
        name   : string;
        fields : 'src field list;
      }

  and 'src field = Field : {
    name        : string;
    output_type : 'out typ;
    resolve     : 'src -> 'out;
  } -> 'src field
end

With these definitions, we can finally construct our GraphQL user type:

(* user Graphql.typ *)
let user_typ = Graphql.Object {
  name = "User";
  fields = [id_field; name_field];
}

The final step is being able to convert a GraphQL type to JSON given a value of the corresponding source type. OCaml generally requires no type annotations due to great type inference capabilities. However, the use of GADTs sometimes necessitates use of type annotations. In the following code snippet, you can read 'a. 'a -> ... as “for all a”:

module Graphql = struct
  (* ... *)

  let rec to_json : 'src. 'src -> 'src typ -> json =
    fun src typ ->
      match typ with
      | Scalar s ->
          s.serialize src
      | Object o ->
          let members = List.map (resolve_field src) o.fields in
          Object members

  and resolve_field : 'src. 'src -> 'src field -> string * json =
    fun src (Field field) ->
      let field_src  = field.resolve src in
      let field_json = to_json field_src field.output_type in
      (field.name, field_json)
end

We’ve reached the crescendo! The ultimate proof is the ability to convert a user to JSON using Graphql.to_json:

let user = { name = "Alice"; id = 1 } in
Graphql.to_json user user_typ
- : json = Object ([("id", Int 1), ("name", String "Alice")])

Victory! Though it could seem like we’ve just replicated the very simple functionality of user_to_json in a more complex manner, what we’ve achieved is much more significant:

These are all topics for future blog posts.

Conclusion and Next Up

This blog post describes a composable, extensible core for a GraphQL server implementation, which allows introspection and ensures to runtime exceptions by capturing complex invariants in the type system. In turn, this means users of the library can enjoy compile-time guarantees and build high-quality GraphQL APIs.

Future blog posts will build more features on top of this core, adding support for:

comments powered by Disqus