This serves as a supplement to Effective Go, based on years of experience and inspiration/ideas from conference talks.
- Add context to errors
- Consistent error and log messages
- Dependency management
- Structured logging
- Avoid global variables
- Testing
- Use linters
- Use goimports
- Use meaningful variable names
- Avoid side effects
- Favour pure functions
- Don't over-interface
- Don't under-package
- Handle signals
- Divide imports
- Avoid unadorned return
- Use canonical import path
- Avoid empty interface
- Main first
- Use internal packages
- Avoid helper/util
- Embed binary data
- Use io.WriteString
- Use functional options
- Structs
- Consistent header naming
Don't:
file, err := os.Open("foo.txt")
if err != nil {
return err
}
Using the approach above can lead to unclear error messages because of missing context.
Do:
file, err := os.Open("foo.txt")
if err != nil {
return fmt.Errorf("open foo.txt failed: %w", err)
}
Wrapping errors with a custom message provides context as it gets propagated up the stack. This does not always make sense. If you're unsure if the context of a returned error is at all times sufficient, wrap it.
Use modules, since it is the built-in go dependency management tooling and will be widely supported (available with Go 1.11+).
Tag your packages using Semantic Versioning, check the modules wiki for more information about
best practices regarding releases.
The git tag for your go package should have the format v<major>.<minor>.<patch>
, e.g., v1.0.1
.
Don't:
log.Printf("Listening on :%d", port)
http.ListenAndServe(fmt.Sprintf(":%d", port), nil)
// 2017/07/29 13:05:50 Listening on :80
Do:
import "github.com/sirupsen/logrus"
// ...
logger.WithField("port", port).Info("Server is listening")
http.ListenAndServe(fmt.Sprintf(":%d", port), nil)
// {"level":"info","msg":"Server is listening","port":"7000","time":"2017-12-24T13:25:31+01:00"}
This is a harmless example, but using structured logging makes debugging and log parsing easier.
Don't:
var db *sql.DB
func main() {
db = // ...
http.HandleFunc("/drop", DropHandler)
// ...
}
func DropHandler(w http.ResponseWriter, r *http.Request) {
db.Exec("DROP DATABASE prod")
}
Global variables make testing and readability hard and every method has access to them (even those, that don't need it).
Do:
func main() {
db := // ...
handlers := Handlers{DB: db}
http.HandleFunc("/drop", handlers.DropHandler)
// ...
}
type Handlers struct {
DB *sql.DB
}
func (h *Handlers) DropHandler(w http.ResponseWriter, r *http.Request) {
h.DB.Exec("DROP DATABASE prod")
}
Use structs to encapsulate the variables and make them available only to those functions that actually need them by making them methods implemented for that struct.
Alternatively higher-order functions can be used to inject dependencies via closures.
func main() {
db := // ...
http.HandleFunc("/drop", DropHandler(db))
// ...
}
func DropHandler(db *sql.DB) http.HandleFunc {
return func (w http.ResponseWriter, r *http.Request) {
db.Exec("DROP DATABASE prod")
}
}
If you really need global variables or constants, e.g., for defining errors or string constants, put them at the top of your file.
Don't:
import "xyz"
func someFunc() {
//...
}
const route = "/some-route"
func someOtherFunc() {
// usage of route
}
var NotFoundErr = errors.New("not found")
func yetAnotherFunc() {
// usage of NotFoundErr
}
Do:
import "xyz"
const route = "/some-route"
var NotFoundErr = errors.New("not found")
func someFunc() {
//...
}
func someOtherFunc() {
// usage of route
}
func yetAnotherFunc() {
// usage of NotFoundErr
}
Don't:
func TestAdd(t *testing.T) {
actual := 2 + 2
expected := 4
if (actual != expected) {
t.Errorf("Expected %d, but got %d", expected, actual)
}
}
Do:
import "github.com/stretchr/testify/assert"
func TestAdd(t *testing.T) {
actual := 2 + 2
expected := 4
assert.Equal(t, expected, actual)
}
Using assert libraries makes your tests more readable, requires less code and provides consistent error output.
Don't:
func TestSomeFunctionSuccess(t *testing.T) {
// ...
}
func TestSomeFunctionWrongInput(t *testing.T) {
// ...
}
Do:
func TestSomeFunction(t *testing.T) {
t.Run("success", func(t *testing.T){
//...
})
t.Run("wrong input", func(t *testing.T){
//...
})
}
Don't:
func TestAdd(t *testing.T) {
assert.Equal(t, 1+1, 2)
assert.Equal(t, 1+-1, 0)
assert.Equal(t, 1, 0, 1)
assert.Equal(t, 0, 0, 0)
}
The above approach looks simpler, but it's much harder to find a failing case, especially when having hundreds of cases.
Do:
func TestAdd(t *testing.T) {
cases := []struct {
A, B, Expected int
}{
{1, 1, 2},
{1, -1, 0},
{1, 0, 1},
{0, 0, 0},
}
for _, tc := range cases {
t.Run(fmt.Sprintf("%d + %d", tc.A, tc.B), func(t *testing.T) {
t.Parallel()
assert.Equal(t, t.Expected, tc.A+tc.B)
})
}
}
Using table driven tests in combination with subtests gives you direct insight about which case is failing and which cases are tested. – Mitchell Hashimoto at GopherCon 2017
Running subtests in parallel allows you to have a lot more test cases and still get those awesomely fast go build times. – The Go Blog
Don't:
func TestRun(t *testing.T) {
mockConn := new(MockConn)
run(mockConn)
}
Do:
func TestRun(t *testing.T) {
ln, err := net.Listen("tcp", "127.0.0.1:0")
t.AssertNil(t, err)
var server net.Conn
go func() {
defer ln.Close()
server, err := ln.Accept()
t.AssertNil(t, err)
}()
client, err := net.Dial("tcp", ln.Addr().String())
t.AssertNil(err)
run(client)
}
Only use mocks if not otherwise possible, favor real implementations. – Mitchell Hashimoto at GopherCon 2017
Don't:
type myType struct {
id int
name string
irrelevant []byte
}
func TestSomething(t *testing.T) {
actual := &myType{/* ... */}
expected := &myType{/* ... */}
assert.True(t, reflect.DeepEqual(expected, actual))
}
Do:
type myType struct {
id int
name string
irrelevant []byte
}
func (m *myType) testString() string {
return fmt.Sprintf("%d.%s", m.id, m.name)
}
func TestSomething(t *testing.T) {
actual := &myType{/* ... */}
expected := &myType{/* ... */}
if actual.testString() != expected.testString() {
t.Errorf("Expected '%s', got '%s'", expected.testString(), actual.testString())
}
// or assert.Equal(t, actual.testString(), expected.testString())
}
Using testString()
for comparing structs helps on complex structs with many
fields that are not relevant for the equality check.
This approach only makes sense for very big or tree-like structs.
– Mitchell Hashimoto at GopherCon 2017
Google open sourced their go-cmp package as a more powerful and safer alternative to reflect.DeepEqual
.
– Joe Tsai.
Only test unexported funcs if you can't access a path via exported funcs. Since they are unexported, they are prone to change.
func ExamleSomeInterface_SomeMethod(){
instance := New()
result, err := instance.SomeMethod()
fmt.Println(result, err)
// Output: someResult, <nil>
}
Use all the linters included in golangci-lint to lint your projects before committing.
# Installation - replace vX.X.X with the version you want to use
GO111MODULE=on go get github.com/golangci/golangci-lint/cmd/[email protected]
# traditional way without go module
go get -u github.com/golangci/golangci-lint/cmd/golangci-lint
# Usage in the project workspace
golangci-lint run
For detailed usage and the ci-pipeline installation guide visit golangci-lint.
Only commit gofmt'd files. Use goimports
for this to format/update the import statements as well.
Avoid single-letter variable names. They may seem more readable to you at the moment of writing but they make the code hard to understand for your colleagues and your future self.
Don't:
func findMax(l []int) int {
m := l[0]
for _, n := range l {
if n > m {
m = n
}
}
return m
}
Do:
func findMax(inputs []int) int {
max := inputs[0]
for _, value := range inputs {
if value > max {
max = value
}
}
return max
}
Single-letter variable names are fine in the following cases.
- They are absolut standard like ...
t
in testsr
andw
in http request handlersi
for the index in a loop
- They name the receiver of a method, e.g.,
func (s *someStruct) myFunction(){}
Of course also too long variables names like createInstanceOfMyStructFromString
should be avoided.
Don't:
func init() {
someStruct.Load()
}
Side effects are only okay in special cases (e.g. parsing flags in a cmd). If you find no other way, rethink and refactor.
In computer programming, a function may be considered a pure function if both of the following statements about the function hold:
- The function always evaluates the same result value given the same argument value(s). The function result value cannot depend on any hidden information or state that may change while program execution proceeds or between different executions of the program, nor can it depend on any external input from I/O devices.
- Evaluation of the result does not cause any semantically observable side effect or output, such as mutation of mutable objects or output to I/O devices.
Don't:
func MarshalAndWrite(some *Thing) error {
b, err := json.Marshal(some)
if err != nil {
return err
}
return ioutil.WriteFile("some.thing", b, 0644)
}
Do:
// Marshal is a pure func (even though useless)
func Marshal(some *Thing) ([]bytes, error) {
return json.Marshal(some)
}
// ...
This is obviously not possible at all times, but trying to make every possible func pure makes code more understandable and improves debugging.
Don't:
type Server interface {
Serve() error
Some() int
Fields() float64
That() string
Are([]byte) error
Not() []string
Necessary() error
}
func debug(srv Server) {
fmt.Println(srv.String())
}
func run(srv Server) {
srv.Serve()
}
Do:
type Server interface {
Serve() error
}
func debug(v fmt.Stringer) {
fmt.Println(v.String())
}
func run(srv Server) {
srv.Serve()
}
Favour small interfaces and only expect the interfaces you need in your funcs.
Deleting or merging packages is far more easier than splitting big ones up. When unsure if a package can be split, do it.
Don't:
func main() {
for {
time.Sleep(1 * time.Second)
ioutil.WriteFile("foo", []byte("bar"), 0644)
}
}
Do:
func main() {
logger := // ...
sc := make(chan os.Signal, 1)
done := make(chan bool)
go func() {
for {
select {
case s := <-sc:
logger.Info("Received signal, stopping application",
zap.String("signal", s.String()))
done <- true
return
default:
time.Sleep(1 * time.Second)
ioutil.WriteFile("foo", []byte("bar"), 0644)
}
}
}()
signal.Notify(sc, os.Interrupt, os.Kill)
<-done // Wait for go-routine
}
Handling signals allows us to gracefully stop our server, close open files and connections and therefore prevent file corruption among other things.
Don't:
import (
"encoding/json"
"github.com/some/external/pkg"
"fmt"
"github.com/this-project/pkg/some-lib"
"os"
)
Do:
import (
"encoding/json"
"fmt"
"os"
"github.com/bahlo/this-project/pkg/some-lib"
"github.com/bahlo/another-project/pkg/some-lib"
"github.com/bahlo/yet-another-project/pkg/some-lib"
"github.com/some/external/pkg"
"github.com/some-other/external/pkg"
)
Divide imports into four groups sorted from internal to external for readability:
- Standard library
- Project internal packages
- Company internal packages
- External packages
Don't:
func run() (n int, err error) {
// ...
return
}
Do:
func run() (n int, err error) {
// ...
return n, err
}
Named returns are good for documentation, unadorned returns are bad for readability and error-prone.
Don't:
package sub
Do:
package sub // import "github.com/my-package/pkg/sth/else/sub"
Adding the canonical import path adds context to the package and makes importing easy.
Don't:
func run(foo interface{}) {
// ...
}
Empty interfaces make code more complex and unclear, avoid them where you can.
Don't:
package main // import "github.com/me/my-project"
func someHelper() int {
// ...
}
func someOtherHelper() string {
// ...
}
func Handler(w http.ResponseWriter, r *http.Reqeust) {
// ...
}
func main() {
// ...
}
Do:
package main // import "github.com/me/my-project"
func main() {
// ...
}
func Handler(w http.ResponseWriter, r *http.Reqeust) {
// ...
}
func someHelper() int {
// ...
}
func someOtherHelper() string {
// ...
}
Putting main()
first makes reading the file a lot more easier. Only the
init()
function should be above it.
If you're creating a cmd, consider moving libraries to internal/
to prevent
import of unstable, changing packages.
Use clear names and try to avoid creating a helper.go
, utils.go
or even
package.
To enable single-binary deployments, use tools to add templates and other static assets to your binary (e.g. github.com/gobuffalo/packr).
A number of important types that satisfy io.Writer
also have a WriteString
method, including *bytes.Buffer
, *os.File
and *bufio.Writer
. WriteString
is behavioral contract with implicit assent that passed string will be written
in efficient way, without a temporary allocation. Therefore using
io.WriteString
may improve performance at most, and at least string will be
written in any way.
Don't:
var w io.Writer = new(bytes.Buffer)
str := "some string"
w.Write([]byte(str))
Do:
var w io.Writer = new(bytes.Buffer)
str := "some string"
io.WriteString(w, str)
func main() {
// ...
startServer(
WithPort(8080),
WithTimeout(1 * time.Second),
)
}
type Config struct {
port int
timeout time.Duration
}
type ServerOpt func(*Config)
func WithPort(port int) ServerOpt {
return func(cfg *Config) {
cfg.port = port
}
}
func WithTimeout(timeout time.Duration) ServerOpt {
return func(cfg *Config) {
cfg.timeout = timeout
}
}
func startServer(opts ...ServerOpt) {
cfg := new(Config)
for _, fn := range opts {
fn(cfg)
}
// ...
}
If a struct has more than one field include field names when instantiating it.
Don't:
params := myStruct{
1,
true,
}
Do:
params := myStruct{
Foo: 1,
Bar: true,
}
Using the normal syntax instead of the new
keyword makes it more clear what is happening: a new instance of the struct is created MyStruct{}
and we get the pointer for it with &
.
Don't:
s := new(MyStruct)
Do:
s := &MyStruct{}
Don't:
r.Header.Get("authorization")
w.Header.Set("Content-type")
w.Header.Set("content-type")
w.Header.Set("content-Type")
Do:
r.Header.Get("Authorization")
w.Header.Set("Content-Type")