/rte

Extraordinary routing

Primary LanguageGoMIT LicenseMIT

rte - routing table extraordinaire

Tests Lint GoDoc Coverage

Simple, opinionated, performant routing.

  • Intuitive, legible interface; encourages treating routing configuration as data to be passed around and manipulated.
  • Extracted path variables are matched to type signature; handlers get to business logic faster, code is more explicit, and programming errors are surfaced before request handling time.
  • Fast AF -- completely avoids the heap during request handling; markedly faster than any other router in the go-http-routing-benchmark suite.
package main

import (
    "fmt"
    "github.com/jwilner/rte"
    "log"
    "net/http"
)

func main() {
    log.Fatal(http.ListenAndServe(":8080", rte.Must(rte.Routes(
        "/foo", rte.Routes(
            "POST", func(w http.ResponseWriter, r *http.Request) {
                // POST /foo
            },
            "/:id", rte.Routes(
                "GET", func(w http.ResponseWriter, r *http.Request, id string) {
                    // GET /foo/:id
                },
                "PUT", func(w http.ResponseWriter, r *http.Request, id string) {
                    // PUT /foo/:id
                },
                "DELETE", func(w http.ResponseWriter, r *http.Request, id string) {
                    // DELETE /foo/:id
                },
                rte.MethodAny, func(w http.ResponseWriter, r *http.Request, id string) {
                    // Match any other HTTP method on /foo/:id (e.g. to serve a 405)
                },
                "POST /bar", func(w http.ResponseWriter, r *http.Request, id string) {
                    // POST /foo/:id/bar
                },
            ),
        ),
    ))))
}

Usage

rte.Route values are passed to rte.Must or rte.New, which constructs an *rte.Table. There are plenty of examples here, but also check out the go docs.

rte.Route

The core value is the simple rte.Route struct:

route := rte.Route{Method: "GET", Path: "/health", Handler: healthHandler}

You can construct and manipulate them as you would any struct; the handler can be a standard http.Handler, http.HandlerFunc, or one of:

func(http.ResponseWriter, *http.Request, string)
func(http.ResponseWriter, *http.Request, string, string)
func(http.ResponseWriter, *http.Request, string, string, string)
func(http.ResponseWriter, *http.Request, [N]string) // where N is a number between 1 and 8

If the handler has string parameters, RTE injects any variables indicted w/in the path into function signature. For signatures of 4 or more, only array signatures are provided; arrays, rather than slices, are used to avoid heap allocations -- and to be explicit. It's a configuration error if the number of path variables doesn't match the number of function parameters (an exception is made for zero string parameter functions -- they can be used with any number of path variables).

Each struct can also be assigned middleware behavior:

route.Middleware = func(w http.ResponseWriter, r *http.Request, next http.Handler) {
    _, _ = fmt.Println(w, "Before request")
    next.ServeHttp(w, r)
    _, _ = fmt.Println(w, "After request")
}

rte.Routes constructor

When it's useful, there's an overloaded variadic constructor:

routes := rte.Routes(
    "GET /health", healthHandler,
    "GET /foo/:foo_id", func(w http.ResponseWriter, r *http.Request, fooID string) {
        _, _ = fmt.Fprintf(w, "fooID: %v", fooID)
    },
)

It can be used to construct hierarchical routes:

routes := rte.Routes(
    "/foo", rte.Routes(
        "POST", func(w http.ResponseWriter, r *http.Request) {
            // POST /foo
        },
        "/:foo_id", rte.Routes(
            "GET", func(w http.ResponseWriter, r *http.Request, fooID string) {
                // GET /foo/:foo_id
            },
            "PUT", func(w http.ResponseWriter, r *http.Request, fooID string) {
                // PUT /foo/:foo_id
            },
        ), myMiddleware
    )
)

The above is exactly equivalent to:

routes := []rte.Route {
    {
        Method: "POST", Path: "/foo",
        Handler: func(http.ResponseWriter, *http.Request) {
            // POST /foo
        },
    },
    {
        Method: "GET", Path: "/foo/:foo_id",
        Handler: func(http.ResponseWriter, *http.Request, id string) {
            // GET /foo/:foo_id
        },
        Middleware: myMiddleware
    },
    {
        Method: "PUT", Path: "/foo/:foo_id",
        Handler: func(w http.ResponseWriter, r *http.Request, id string) {
            // PUT /foo/:foo_id
        },
        Middleware: myMiddleware
    },
}

See examples_test.go or go docs for more examples.

Compiling the routing table

Zero or more routes are combined to create a Table handler using either New or Must:

var routes []rte.Route
tbl, err := rte.New(routes) // errors on misconfiguration
tbl = rte.Must(routes) // panics on misconfiguration

If you're dynamically constructing your routes, the returned rte.Error type helps you recover from misconfiguration.

*rte.Table satisfies the standard http.Handler interface and can be used with standard Go http utilities.

Extras

RTE provides a few basic values and functions to help with common patterns. Many of these functions take in a []Route and return a new, potentially modified []Route, in keeping with the design principles.

MethodAny

RTE performs wildcard matching in paths with the : syntax; it can also perform wildcard matching of methods via the use of rte.MethodAny. You can use rte.MethodAny anywhere you would a normal HTTP method; it will match any requests to the path that don't match an explicit, static method:

rte.Routes(
    // handles GETs to /
    "GET /", func(http.ResponseWriter, *http.Request){},
    // handles POST PUT, OPTIONS, etc. to /
    rte.MethodAny+" /", func(http.ResponseWriter, *http.Request){},
)

DefaultMethod

rte.DefaultMethod adds a rte.MethodAny handler to every path; useful if you want to serve 405s for all routes.

reflect.DeepEqual(
    rte.DefaultMethod(
        hndlr405,
        []rte.Route {
            {Method: "GET", Path: "/foo", Handler: fooHandler},
            {Method: "POST", Path: "/foo", Handler: postFooHandler},
            {Method: "GET", Path: "/bar", Handler: barHandler},
        },
    ),
    []rte.Route {
        {Method: "GET", Path: "/foo", Handler: fooHandler},
        {Method: rte.MethodAny, Path: "/foo", Handler: hndlr405},
        {Method: "POST", Path: "/foo", Handler: postFooHandler},
        {Method: "GET", Path: "/bar", Handler: barHandler},
        {Method: rte.MethodAny, Path: "/bar", Handler: hnldr405},
    },
)

Wrap

rte.Wrap adds middleware behavior to every contained path; if a middleware is already set, the new middleware will be wrapped around it -- so that the stack will have the new middleware at the top, the old middleware in the middle, and the handler at the bottom.

reflect.DeepEqual(
    rte.Wrap(
        myMiddleware,
        []rte.Route {
            {Method: "GET", Path: "/foo", Handler: myHandler},
        },
    ),
    []rte.Route {
        {Method: "GET", Path: "/foo", Handler: myHandler, Middleware: myMiddleware},
    },
)

OptTrailingSlash

OptTrailingSlash makes each handler also match its slashed or not-slashed version.

reflect.DeepEqual(
    rte.OptTrailingSlash(
        []rte.Route {
            {Method: "GET", Path: "/foo", Handler: myHandler},
            {Method: "POST", Path: "/bar/", Handler: barHandler},
        },
    ),
    []rte.Route {
        {Method: "GET", Path: "/foo", Handler: myHandler},
        {Method: "GET", Path: "/foo/", Handler: myHandler},
        {Method: "POST", Path: "/bar/", Handler: barHandler},
        {Method: "POST", Path: "/bar", Handler: barHandler},
    },
)

more

Check out the go docs for still more extras.

Trade-offs

It's important to note that RTE uses a fixed size array of strings for path variables paired with generated code to avoid heap allocations; currently, this number is fixed at 8, which means that RTE does not support routes with more than eight path variables (doing so will cause an error or panic). The author is deeply skeptical that anyone actually really needs more than eight path variables; that said, it's a design goal to provide support for higher numbers once the right packaging technique is found.

Performance

Modern Go routers place a lot of emphasis on speed. There's plenty of room for skepticism about this attitude, as most web application will be IO bound. Nonetheless, this is the barrier for entry to the space these days. To this end, RTE completely avoids performing zero heap allocations while serving requests and uses a relatively optimized data structure (a compressed trie) to route requests.

Benchmarks are drawn from this fork of go-http-routing-benchmark (which appears unmaintained). The benchmarks were run on a 2013 MB Pro with a 2.6 GHz i5 and 8 GB ram. For fun, RTE is compared to some of the most popular Go router layers below.

Single Param Micro Benchmark Reps ns/op B/op allocs/op
RTE 20000000 64.0 0 0
Gin 20000000 79.4 0 0
Echo 20000000 105 0 0
HttpRouter 10000000 138 32 1
Beego 1000000 1744 352 3
GorillaMux 300000 3775 1280 10
Martini 200000 6891 1072 10
Five Param Micro Benchmark Reps ns/op B/op allocs/op
RTE 20000000 116 0 0
Gin 10000000 137 0 0
Echo 5000000 253 0 0
HttpRouter 3000000 416 160 1
Beego 1000000 2036 352 3
GorillaMux 300000 5194 1344 10
Martini 200000 8091 1232 11
Github API with 1 Param Reps ns/op B/op allocs/op
RTE 10000000 156 0 0
Gin 10000000 184 0 0
Echo 5000000 266 0 0
HttpRouter 5000000 296 96 1
Beego 1000000 2018 352 3
GorillaMux 200000 10949 1296 10
Martini 100000 14957 1152 11
Google Plus API with 1 Param Reps ns/op B/op allocs/op
RTE 20000000 89.1 0 0
Gin 20000000 123 0 0
Echo 10000000 143 0 0
HttpRouter 10000000 185 64 1
Beego 1000000 1488 352 3
GorillaMux 300000 4053 1280 10
Martini 200000 6375 1072 10
Google Plus API with 2 Params Reps ns/op B/op allocs/op
RTE 10000000 139 0 0
Gin 10000000 141 0 0
HttpRouter 5000000 225 64 1
Echo 10000000 237 0 0
Beego 1000000 1646 352 3
GorillaMux 200000 8675 1296 10
Martini 100000 13586 1200 13

Design principles

simplicity

RTE attempts to follow Golang's lead by avoiding features which complicate routing behavior when a viable alternative is available to the user. For example, RTE chooses not to allow users to specify path variables with types -- e.g. {foo_id:int} -- or catch-all paths -- e.g. a /foo/* matching /foo/bar/blah. Supporting either of those features would introduce complicated precedence behavior, and simple alternatives exist for users.

Additionally, RTE aims to have defined behavior in all circumstances and to document and prove that behavior with unit tests.

limiting state space

Many modern routers coordinate by mutating a common data structure -- unsurprisingly, usually called a Router. In larger applications that pass around the routers and subrouters, setting different flags and options in different locations, the state of the router can be at best hard to reason about and at worst undefined -- it is not uncommon for certain feature combinations to fail or have unexpected results.

By centering on the simple rte.Route and not exposing any mutable state, RTE keeps its interface simple to understand, while also simplifying its own internals. Because most routing libraries focus on the mutable router object, they do not have explicit finalization, and thus their internal logic must remain open to modification at any time -- RTE does not have this problem.

When a routing feature is necessary, it will usually be added as an helper method orthogonal to the rest of the API. For example, rather than providing a method OptTrailingSlash(enabled bool) on *rte.Table and pushing complexity into the routing logic, RTE provides the pure function rte.OptTrailingSlash(routes []rte.Route) []rte.Route, which adds the new rte.Routes necessary to optionally match trailing slashes, while the routing logic remains unchanged.

reflect.DeepEqual(
    rte.OptTrailingSlash(
        []Route{
            {Method: "GET", Path: "/foo", Handler: myHandler},
        },
    ),
    []Route {
        {Method: "GET", Path: "/foo", Handler: myHandler},
        {Method: "GET", Path: "/foo/", Handler: myHandler},
    }
)

Development

Check out the Makefile for dev entrypoints.

TLDR:

  • make test
  • make test-cover
  • make gen (regenerates internal code)
  • make check (requires golint -- install with go get -u golang.org/x/lint/golint)

CI

Travis builds. In addition to tests, the build is gated on golint and whether the checked-in generated code matches the code as currently generated.