Skip to content

Latest commit

 

History

History
767 lines (621 loc) · 25.9 KB

README.md

File metadata and controls

767 lines (621 loc) · 25.9 KB

Build Status

A GraphQL Server library - in Erlang

This project contains the necessary support code to implement GraphQL servers in Erlang. Its major use is on top of some other existing transport library, for instance the cowboy web server. When a request arrives, it can be processed by the GraphQL support library and a GraphQL answer can be given. In a way, this replaces all of your REST endpoints with a single endpoint: one for Graph Queries.

This README provides the system overview and its mode of operation.

Changelog

See the file CHANGELOG.md in the root of the project.

Status

Currently, the code implements all of the October 2016 GraphQL specification, except for a few areas:

  • Some validators are missing and pending implementation. The important validators are present, however. Missing stuff are all tracked as issues in this repository.
  • Parametrization inside fragments are not yet implemented fully.

In addition, we are working towards June 2018 compliance. We already implemented many of the changes in the system. But we are still missing some parts. The implementation plan is on a demand driven basis for Shopgun currently, in that we tend to implement things when there is a need for them.

Documentation

This is a big library. In order to ease development, we have provided a complete tutorial for GraphQL Erlang:

https://github.com/shopgun/graphql-erlang-tutorial

Also, the tutorial has a book which describes how the tutorial example is implemented in detail:

https://shopgun.github.io/graphql-erlang-tutorial/

NOTE: Read the tutorial before reading on in this repository if you haven't already. This README gives a very quick overview, but the canonical documentation is the book at the moment.

What is GraphQL

GraphQL is a query language for the web. It allows a client to tell the server what it wants in a declarative way. The server then materializes a response based on the clients query. This makes your development client-centric and client-driven, which tend to be a lot faster from a development perspective. A project is usually driven from the top-of-the-iceberg and down, so shuffling more onus on the client side is a wise move in modern system design.

GraphQL is also a contract. Queries and responses are typed and contract-verified on both the input and output side. That is, GraphQL also acts as a contract-checker. This ensures:

  • No client can provide illegal queries to the server backend. These are filtered out by the GraphQL layer.
  • No server can provide illegal responses to the client. These are altered such that the client gets a valid response according to the schema by replacing failing nodes with null-values.
  • The contract documents the API
  • The contract describes how the client can query data. This is the closest to HATEOAS we will probably get without going there.
  • Queries tend to be large and all-encompassing. This means we don't pay the round-trip-time for a request/response like you do in e.g., HTTP and HTTP/2 based systems where multiple queries are executed back to back and depends on each other. Almost every query can be handled in a single round trip.

Finally, GraphQL supports introspection of its endpoint. This allows systems to query the server in order to learn what the schema is. In turn, tooling can be built on top of GraphQL servers to provide development-debug user interfaces. Also, languages with static types can use the introspection to derive a type model in code which matches the contract. Either by static code generation, or by type providers.

Whirlwind tour

The GraphQL world specifies a typed schema definition. For instance the following taken from the Relay Modern specification:

interface Node {
  id: ID!
}

type Faction : Node {
  id: ID!
  name: String
  ships: ShipConnection
}

type Ship : Node {
  id: ID!
  name: String
}

type ShipConnection {
  edges: [ShipEdge]
  pageInfo: PageInfo!
}

type ShipEdge {
  cursor: String!
  node: Ship
}

type PageInfo {
  hasNextPage: Boolean!
  hasPreviousPage: Boolean!
  startCursor: String
  endCursor: String
}

type Query {
  rebels: Faction
  empire: Faction
  node(id: ID!): Node
}

input IntroduceShipInput {
  factionId: String!
  shipNamed: String!
  clientMutationId: String!
}

type IntroduceShipPayload {
  faction: Faction
  ship: Ship
  clientMutationId: String!
}

type Mutation {
  introduceShip(input: IntroduceShipInput!): IntroduceShipPayload
}

The schema is a subset of the Star Wars schema given as the typical GraphQL example all over the web. The GraphQL world roughly splits the world into input objects and output objects. Input objects are given as part of a query request by the client. Output objects are sent back from the server to the client.

This Erlang implementation contains a schema parser for schemas like the above. Once parsed, a mapping is provided by the programmer which maps an output type in the schema to an Erlang module. This module must implement a function

-spec execute(Context, Object, Field, Args) ->
    {ok, Response}
  | {error, Reason}.

which is used to materialize said object. That is, when you request a field F in the object O, a call is made to execute(Context, O, F, Args). The value Context provides a global context for the query. It is used for authentication data, for origin IP addresses and so on. The context is extensible by the developer with any field they need. The Args provides arguments for the field. Look, for instance at the type Mutation and the introduceShip field, which takes an argument input of type IntroduceShipInput!.

Materialization is thus simply a function call in the Erlang world. These calls tend to be used in two ways: Either they acquire a piece of data from a database (e.g., mnesia) and return that data as an Object. Or they materialize fields on an already loaded object. When execution of a query is processed, you can imagine having a "cursor" which is being moved around in the result set and is used to materialize each part of the query.

For example, look at the following query:

query Q {
  node(id: "12098141") {
      ... on Ship {
        id
        name
      }
  }
}

When this query executes, it will start by a developer provided initial object. Typically the empty map #{}. Since the node field is requested, a call is performed to match:

-module(query).

...
execute(Ctx, #{}, <<"node">>, #{ <<"id">> := ID }) ->
    {ok, Obj} = load_object(ID).

Now, since you are requesting the id and name fields on a Ship inside the node, the system will make a callback to a type-resolver for the Obj in order to determine what type it is. We omit that part here, but if it was something else, a faction say, then the rest of the query would not trigger. Once we know that id "12098141" is a Ship, we "move the cursor" to a ship and calls the execute function there:

-module(ship).

-record(ship, { id, name }).

execute(Ctx, #ship{ id = Id }, <<"id">>, _Args) ->
    {ok, ID};
execute(Ctx, #ship{ name = Name }, <<"name">>, _Args) ->
    {ok, Name}.

Two materialization calls will be made. One for the field <<"id">> and one for the field <<"name">>. The end result is then materialized as a response to the caller.

Materilization through derivation

A common use of the functions is to derive data from existing data. Suppose we extend the ship in the following way:

type Ship {
  ...
  capacity : float!
  load : float!
  loadRatio : float!
}

so a ship has a certain capacity and a current load in its cargo bay. We could store the loadRatio in the database and keep it up to date. But a more efficient way to handle this is to compute it from other data:

-module(ship).

-record(ship,
    { id,
      name,
      capacity,
      load }).

execute(...) ->
  ...;
execute(Ctx, #ship {
                capacity = Cap,
                load = Load }, <<"loadRatio">>, _) ->
    {ok, Load / Cap };
...

This will compute that field if it is requested, but not compute it when it is not requested by a client. Many fields in a data set are derivable in this fashion. Especially when a schema changes and grows over time. Old fields can be derived for backwards compatibility and new fields can be added next to it.

In addition, it tends to be more efficient. A sizable portion of modern web work is about moving data around. If you have to move less data, you decrease the memory and network pressure, which can translate to faster service.

Materializing JOINs

If we take a look at the Faction type, we see the following:

type Faction : Node {
  id: ID!
  name: String
  ships: ShipConnection
}

in this, ships is a field referring to a ShipConnection. A Connection type is Relay Modern standard of how to handle a paginated set of objects in GraphQL. Like "Materialization by derivation" we would derive this field by looking up the data in the database for the join and then producing an object which the ship_connection_resource can handle. For instance:

execute(Ctx, #faction { id = ID }, <<"ships">>, _Args) ->
    {ok, Ships} = ship:lookup_by_faction(ID),
    pagination:build_pagination(Ships).

where the build_pagination function returns some object which is a generic connection object. It will probably look something along the lines of

#{
  '$type' => <<"ShipConnection">>,
  <<"pageInfo">> => #{
      <<"hasNextPage">> => false,
      ...
  },
  <<"edges">> => [
      #{ <<"cursor">> => base64:encode(<<"edge:1">>),
          <<"node">> => #ship{ ... } },
      ...]
}

which can then be processed further by other resources. Note how we are eagerly constructing several objects at once and then exploiting the cursor moves of the GraphQL system to materialize the fields which the client requests. The alternative is to lazily construct materializations on demand, but when data is readily available anyway, it is often more efficient to just pass pointers along.

API

The GraphQL API is defined in the module graphql. Every functionality is exported in that module. Do not call inside other modules as their functionality can change at any point in time even between major releases.

The system deliberately splits each phase and hands it over to the programmer. This allows you to debug a bit easier and gives the programmer more control over the parts. A typical implementation will start by using the schema loader:

inject() ->
  {ok, File} = application:get_env(myapp, schema_file),
  Priv = code:priv_dir(myapp),
  FName = filename:join([Priv, File]),
  {ok, SchemaData} = file:read_file(FName),
  Map = #{
    scalars => #{ default => scalar_resource },
    interfaces => #{ default => resolve_resource },
    unions => #{ default => resolve_resource },
    objects => #{
      'Ship' => ship_resource,
      'Faction' => faction_resource,
      ...
      'Query' => query_resource,
      'Mutation' => mutation_resource
    }
  },
  ok = graphql:load_schema(Map, SchemaData),
      Root = {root,
      #{
        query => 'Query',
        mutation => 'Mutation',
        interfaces => []
      }},
  ok = graphql:insert_schema_definition(Root),
  ok = graphql:validate_schema(),
  ok.

This will set up the schema in the code by reading it from a file on disk. Each of the _resource names refers to modules which implements the backend code.

In order to execute queries on the schema, code such as the following can be used. We have a query document in Doc and we have a requested operation name in OpName and parameter variables for the given op in Vars. The variables Req and State are standard cowboy request and state tracking variables from cowboy_rest.

run(Doc, OpName, Vars, Req, State) ->
  case graphql:parse(Doc) of
    {ok, AST} ->
      try
          {ok, #{fun_env := FunEnv,
                ast := AST2 }} = graphql:type_check(AST),
          ok = graphql:validate(AST2),
          Coerced = graphql:type_check_params(FunEnv, OpName, Vars),
          Ctx = #{ params => Coerced, operation_name => OpName },
          Response = graphql:execute(Ctx, AST2),
          Req2 = cowboy_req:set_resp_body(encode_json(Response), Req),
          {ok, Reply} = cowboy_req:reply(200, Req2),
          {halt, Reply, State}
      catch
            throw:Err ->
                err(400, Err, Req, State)
      end;
    {error, Error} ->
        err(400, {parser_error, Error}, Req, State)
  end.

Conventions

In this GraphQL implementation, the default value for keys are type binary(). This choice is deliberate, since it makes the code more resistent to atom() overflow and also avoids some conversions between binary() and atom() values in the system. A later version of the library might redesign this aspect, but we are somewhat stuck with it for now.

However, there are many places where you can input atom values and then have them converted internally by the library into binary values. This greatly simplifies a large number of data entry tasks for the programmer. The general rules are:

  • If you supply a value to the system and it is an atom(), the internal representation is a binary value.
  • If the system hands you a value, it is a binary() value and not an atom().

Middlewares

This GraphQL system does not support middlewares, because it turns out the systems design is flexible enough middlewares can be implemented by developers themselves. The observation is that any query runs through the Query type and thus a query_resource. Likewise, any Mutation factors through the mutation_resource.

As a result, you can implement middlewares by using the execute/4 function as a wrapper. For instance you could define a mutation function as:

execute(Ctx, Obj, Field, Args) ->
    AnnotCtx = perform_authentication(Ctx),
    execute_field(AnnotCtx, Obj, Field, Args).

The reason this works so well is because we are able to use pattern matching on execute/4 functions and then specialize them. If we had an individual function for each field, then we would have been forced to implement middlewares in the system, which incurs more code lines to support.

More complex systems will define a stack of middlewares in the list and run them one by one. As an example, a clientMutationId is part of the Relay Modern specification and must be present in every mutation. You can build your mutation_resource such that it runs a maps:take/2 on the argument input, runs the underlying mutation, and then adds back the clientMutationId afterwards.

Schema Definitions

This GraphQL implementation follows the Jun2018 specification for defining a schema. In this format, one writes the schema according to specification, including doc-strings. What was represented as tags in an earlier implementation of GraphQL for Erlang is now represented as a @directive annotation, as per the specification.

As an example, you can write something along the lines of:

"""
A Ship from the Star Wars universe
"""
type Ship : Node {
  "Unique identity of the ship"
  id: ID!

  "The name of the ship"
  name: String
}

And the schema parser knows how to transform this into documentation for introspection.

Resource modules

The following section documents the layout of resource modules as they are used in GraphQL, and what they are needed for in the implementation.

Scalar Resources

GraphQL contains two major kinds of data: objects and scalars. Objects are product types where each element in the product is a field. Raw data are represented as scalar values. GraphQL defines a number of standard scalar values: boolean, integers, floating point numbers, enumerations, strings, identifiers and so on. But you can extend the set of scalars yourself. The spec will contain something along the lines of

scalar Color
scalar DateTime

and so on. These are mapped onto resource modules handling scalars. It is often enough to provide a default scalar module in the mapping and then implement two functions to handle the scalars:

-module(scalar_resource).

-export(
  [input/2,
    output/2]).

-spec input(Type, Value) -> {ok, Coerced} | {error, Reason}
  when
    Type :: binary(),
    Value :: binary(),
    Coerced :: any(),
    Reason :: term().
input(<<"Color">>, C) -> color:coerce(C);
input(<<"DateTime">>, DT) -> datetime:coerce(DT);
input(Ty, V) ->
    error_logger:info_report({coercing_generic_scalar, Ty, V}),
    {ok, V}.

-spec output(Type, Value) -> {ok, Coerced} | {error, Reason}
  when
    Type :: binary(),
    Value :: binary(),
    Coerced :: any(),
    Reason :: term().
output(<<"Color">>, C) -> color:as_binary(C);
output(<<"DateTime">>, DT) -> datetime:as_binary(DT);
output(Ty, V) ->
    error_logger:info_report({output_generic_scalar, Ty, V}),
    {ok, V}.

Scalar Mappings allow you to have an internal and external representation of values. You could for instance read a color such as #aabbcc, convert it into #{ r => 0.66, g => 0.73, b => 0.8 } internally and back again when outputting it. Likewise a datetime object can be converted to a UNIX timestamp and a timezone internally if you want. You can also handle multiple different ways of coercing input data, or have multiple internal data representations.

Type resolution Resources

For GraphQL to function correctly, we must be able to resolve types of concrete objects. This is because the GraphQL system allows you to specify abstract interfaces and unions. An example from the above schema is the Node interface which is implemented by Ship and Faction among other things. If we are trying to materialize a node, the GraphQL must have a way to figure out the type of the object it is materializing. This is handled by the type resolution mapping:

-module(resolve_resource).

-export([execute/1]).

%% The following is probably included from a header file in a real
%% implementation
-record(ship, {id, name}).
-record(faction, {id, name}).

execute(#ship{}) -> {ok, <<"Ship">>};
execute(#faction{}) -> {ok, <<"Faction">>};
execute(Obj) ->
    {error, unknown_type}.

Output object Resources

Each (output) object is mapped onto an Erlang module responsible for handling field requests in that object. The module looks like:

-module(object_resource).

-export([execute/4]).

execute(Ctx, SrcObj, <<"f">>, Args) ->
    {ok, 42};
execute(Ctx, SrcObj, Field, Args) ->
    default

The only function which is needed is the execute/4 function which is called by the system whenever a field is requested in that object. The 4 parameters are as follows:

  • Ctx - The context of the query. It contains information pertaining to the current position in the Graph, as well as user-supplied information from the start of the request. It is commonly used as a read-only store for authentication/authorization data, so you can limit what certain users can see.
  • SrcObj - The current object on which we are operating. Imagine we have two ships, a B-wing and an X-wing. Even if we request the same fields on the two ships, the SrcObj is going to be different. GraphQL often proceeds by having certain fields fetch objects out of a backing store and then moving the cursor onto that object and calling the correct object resource for that type. The SrcObj is set to point to the object that is currently being operated upon.
  • Field - The field in the object which is requested.
  • Args - A map of field arguments. See the next section.

Field Argument rules

In GraphQL, field arguments follow a specific pattern:

  • Clients has no way to input a null value. The only thing they can do is to omit a given field in the input. In particular, clients must supply a field which is non-null.
  • Servers always see every field in the input, even if the client doesn't supply it. If the client does not supply a field, and it has no default value, the server sees a null value for that field.

This pattern means there is a clear way for the client to specify "no value" and a clear way for the server to work with the case where the client specified "no value. It eliminates corner cases where you have to figure out what the client meant.

Resolution follows a rather simple pattern in GraphQL. If a client omits a field and it has a default value, the default value is input. Otherwise null is input. Clients must supply every non-null field.

Note: This limitation is lifted in the Jun2018 GraphQL specification, but this server doesn't implement that detail yet.

On the server side, we handle arguments by supplying a map of KV pairs to the execute function. Suppose we have an input such as

input Point {
    x = 4.0 float
    y float
}

The server can handle this input by matching directly:

execute(Ctx, SrcObj, Field,
    #{ <<"x">> := XVal, <<"y">> := YVal }) ->
  ...

This will always match. If the client provides the input {} which is the empty input, XVal will be 4.0 due to the default value. And YVal will be null. If the client supplies, e.g., { x: 2.0, y: 7.0 } the map #{ <<"x">> => 2.0, <<"y">> => 7.0 } will be provided.

Tips & Tricks

The execute function allows you to make object-level generic handling of fields. If, for example, your SrcObj is a map, you can do generic lookups by using the following handler:

execute(_Ctx, Obj, Field, _Args) ->
    case maps:get(Field, Obj, not_found) of
      not_found -> {ok, null};
      Val -> {ok, Val}
    end.

Another trick is to use generic execution to handle "middlewares" - See the appropriate section on Middlewares.

System Architecture

Most other GraphQL servers provide no type->module mapping. Rather, they rely on binding of individual functions to fields. The implementation began with the same setup, but it turns out pattern matching is a good fit for the notion of requesting different fields inside an object. Thus, we use pattern matching as a destructuring mechanism for incoming queries.

Schema

Internally, the system parses the schema into an ETS table, on which it can perform queries in parallel to satisfy multiple requests at the same time.

A schema injector allows the developer to parse a schema from a file or from memory, then bind exeuction modules to the schemas types. Once finishes, the schema is finalized which runs a lint check over the schema and rejects schemas which are nonsensical.

Query

A query is treated as a compiler chain, which is a design that fits Erlang well. Compilers rely a lot on pattern matching, so we can process a query symbolically by matching on it and gradually transforming it into a query plan which can then be executed.

  • A lexer tokenizes the query
  • A parser constructs an AST of the query from the token stream
  • An type checker walks the AST and attaches type information to the AST by looking up data in the schema ETS table. The pass also detects type and validation errors. The type checker is written in a bi-directional style so it flips between inference, in which we deduce the type of a term, and checking in which we verify a term has a given type.
  • A validator performs additional linting. Many queries are type-correct and thus executable, but are still malformed because they have nonsensical parts in them. The validator phase rejects such queries.
  • A query plan is formed from the AST.
  • An executor runs the query plan.

Of these tasks, only the execution phase in the end is performance-critical. Clients can pre-load query documents to the server, which means the document acts as a stored procedure on the server side. The server can then do parsing, elaboration, type checking and validation once and for all at load time. In addition it provides a security measure: clients in production can only call a pre-validated set of queries if such desired.

User Interface

GraphQL has some very neat Javascript tooling which plugs into the introspection of a GraphQL server and provides additional functionality:

  • GraphiQL - Provides a query interface with autocompletion, documentation, debugging, ways to execute queries and so on. It is highly recommended you add such a system in staging and production as it is indispensable if you are trying to figure out a new query or why a given query returned a specific kind of error.

Additionally, Relay Modern provides specifications for cache refreshing, pagination, mutation specifications and so on. It is recommended you implement those parts in your system as it is part of a de-facto standard for how GraphQL servers tend to operate.

Tests

The GraphQL project has an extensive test suite. We prefer adding regressions to the suite as we experience them. Some of the tests are taken from the the official GraphQL repository and translated. More work is definitely needed, but in general new functionality should be provided together with a test case that demonstrates the new functionality.

The general tests are:

  • dungeon_SUITE which implements a "MUD" style dungeon backend. It is used as a way to handle most of the test cases we cook up ourselves. It is driven by a schema and uses a query document for its queries. If you don't know where to add a test, this is a good place.
  • enum_SUITE Taken from the official Node.js de-facto implementation, this suite uses the "colors" schema in order to verify certain hard-to-get-right properties about enumerated data types.
  • graphql_SUITE Main suite for things which doesn't fit elsewhere. Checks lexing/parsing, "Hello World" style queries and also introspection.
  • star_wars_SUITE An implementation from the specification of the Star Wars universe. Allows us to verify queries from the specification in our own code. Also part of the Node.js de-facto implementation, so it is easy for us to transplant a test from there to here.
  • validation_SUITE GraphQL contains a lot of different validation checks. We handle some of these in the type checker and some in a validation pass. The tests here are mostly verifying parts of the specification. It uses the "Pet" schema as a base.