api-fu (noun)
- (informal) Mastery of APIs. 💪
- The top level
apifu
package is an opinionated library that aims to make it as easy as possible to build APIs that conform to API-fu's ideals. See the examples directory for example usage. - The
graphql
package is an unopinionated library for building GraphQL APIs. If you agree with API-fu's ideals, you should useapifu
instead, but if you want something lower level, thegraphql
package is still an excellent standalone GraphQL library. It fully supports all features of the June 2018 spec. - The
graphql/transport
directory contains unopinionated libraries for using various transport protocols. For example, it contains transport implementations that allow you to serve your GraphQL API via WebSockets and provide subscription functionality. - The
cmd/gql-client-gen
package provides a CLI tool that can be used to generate types for use by client code. THis allows your client's queries to be type-safe and fully validated at compile time.
The above packages are mature and have been thoroughly proven in real-world, production deployments. The following packages have not yet seen such rigorous real-world testing and are thus considered experimental. They are fully functional and well unit tested, but may change at any time and are not yet subject to any compatibility guarantees.
- The
jsonapi
package is a library for building JSON:API APIs. It's somewhat high level, but is no more opinionated than JSON:API itself is. However, it does hold some of those opinions more strongly (i.e. it doesn't support violating many of the JSON:API spec's recommendations and "SHOULD"s).
API-fu builds GraphQL APIs with code. To begin, you need a config that at least defines a query field:
var fuCfg apifu.Config
fuCfg.AddQueryField("foo", &graphql.FieldDefinition{
Type: graphql.StringType,
Resolve: func(ctx *graphql.FieldContext) (interface{}, error) {
return "bar", nil
},
})
From there, you can build the API:
fu, err := apifu.NewAPI(&fuCfg)
if err != nil {
panic(err)
}
And serve it:
fu.ServeGraphQL(w, r)
API-fu also has first-class support for common patterns such as nodes that are queryable using global ids. See the examples directory for more complete example code.
✅ Supports all features of the latest GraphQL spec.
This includes null literals, error extensions, subscriptions, and directives.
The graphql package is over twice as fast and several times more memory efficient than its inspiration (graphql-go/graphql).
pkg: github.com/ccbrown/api-fu/graphql/benchmarks
BenchmarkAPIFu
BenchmarkAPIFu-16 765 1553517 ns/op 890575 B/op 22587 allocs/op
BenchmarkGraphQLGo
BenchmarkGraphQLGo-16 315 3753681 ns/op 3990220 B/op 45952 allocs/op
The graphql
package supports virtually any batching or concurrency pattern using low level primitives.
The apifu
package provides high level ways to use them.
For example, you can define a resolver like this to do work in a goroutine:
fuCfg.AddQueryField("myField", &graphql.FieldDefinition{
Type: graphql.IntType,
Resolve: func(ctx *graphql.FieldContext) (interface{}, error) {
return Go(ctx.Context, func() (interface{}, error) {
return doSomethingComplex(), nil
}), nil
},
})
Or you can define a resolver like this to batch up queries, allowing you to minimize round trips to your database:
fuCfg.AddQueryField("myField", &graphql.FieldDefinition{
Type: graphql.IntType,
Resolve: Batch(func(ctx []*graphql.FieldContext) []graphql.ResolveResult {
return resolveABunchOfTheseAtOnce(ctx)
},
})
For example, the apifu
package provides date-time and long (but JavaScript safe) integers.
📡 Implements handlers for HTTP, the Apollo graphql-ws protocol, and the newer graphql-transport-ws protocol.
Once you've built your API, all you have to do is:
fu.ServeGraphQL(w, r)
Or:
fu.ServeGraphQLWS(w, r)
📖 Provides easy-to-use helpers for creating connections adhering to the Relay Cursor Connections Specification.
Just provide a name, cursor constructor, edge fields, and edge getter:
{
"messagesConnection": apifu.TimeBasedConnection(&apifu.TimeBasedConnectionConfig{
NamePrefix: "ChannelMessages",
EdgeCursor: func(edge interface{}) apifu.TimeBasedCursor {
message := edge.(*model.Message)
return apifu.NewTimeBasedCursor(message.Time, string(message.Id))
},
EdgeFields: map[string]*graphql.FieldDefinition{
"node": &graphql.FieldDefinition{
Type: graphql.NewNonNullType(messageType),
Resolve: func(ctx *graphql.FieldContext) (interface{}, error) {
return ctx.Object, nil
},
},
},
EdgeGetter: func(ctx *graphql.FieldContext, minTime time.Time, maxTime time.Time, limit int) (interface{}, error) {
return ctxSession(ctx.Context).GetMessagesByChannelIdAndTimeRange(ctx.Object.(*model.Channel).Id, minTime, maxTime, limit)
},
}),
}
The gql-client-gen
tool can be used to generate types for use in client-side code as well as validate queries at compile-time. The generated types intelligently unmarshal inline fragments and fragment spreads based on __typename
values.
See cmd/gql-client-gen for details.
During validation, you can specify a max operation cost or get the actual cost of an operation using customizable cost definitions:
doc, errs := graphql.ParseAndValidate(req.Query, req.Schema, req.ValidateCost(maxCost, &actualCost))
The following are guidelines that are recommended for all new GraphQL APIs. API-fu aims to make it easy to conform to these for robust and future-proof APIs:
- All mutations should resolve to result types. No mutations should simply resolve to a node. For example, a
createUser
mutation should resolve to aCreateUserResult
object with auser
field rather than simply resolving to aUser
. This is necessary to keep mutations extensible. Likewise, subscriptions should not resolve directly to node types. For example, a subscription for messages in a chat room (chatRoomMessages
) should resolve to aChatRoomMessagesEvent
type. - Nodes with 1-to-many relationships should make related nodes available via Relay Cursor Connections. Nodes should not have fields that simply resolve to lists of related nodes. Additionally, all connections must require a
first
orlast
argument that specifies the upper bound on the number of nodes returned by that connection. This makes it possible to determine an upper bound on the number of nodes returned by a query before that query begins execution, e.g. using rules similar to GitHub's. - Mutations that modify nodes should always include the updated version of that node in the result. This makes it easy for clients to maintain up-to-date state and tolerate eventual consistency (If a client updates a resource, then immediately requests it in a subsequent query, the server may provide a version of the resource that was cached before the update.).
- Nodes should provide revision numbers. Each time a node is modified, the revision number must increment. This helps clients maintain up-to-date state and enables simultaneous change detection.
- It should be easy for clients to query historical data and subscribe to real-time data without missing anything due to race conditions. The most transparent and fool-proof way to facilitate this is to make subscriptions immediately push a small history of events to clients as soon as they're started. The pushed history should generally only need to cover a few seconds' worth of events. If queries use eventual consistency, the pushed history should be at least as large as the query cache's TTL.
This library is not versioned. However, one guarantee is made: Unless otherwise noted, backwards-incompatible changes made will break your build at compile-time. If your application compiles after updating API-fu, you're good to go.