~

It is hard to write bootstrap tool to quickly create Go service. So I write this guide instead. This is a quick checklist for me every damn time I need to write a Go service from scratch. Also, this is my personal opinion, so feel free to comment.

main.go
internal
    business
        http
            handler.go
            service.go
            models.go

        grpc
            handler.go
            models.go
        consumer
            handler.go
            service.go
            models.go
        service.go
        repository.go
        models.go

All business codes are inside internal. Each business has a different directory business.

Inside each business, there are 2 handlers: http, grpc:

• http is for public APIs (Android, iOS, ... are clients).
• grpc is for internal APIs (other services are clients).
• consumer is for consuming messages from queue (Kafka, RabbitMQ, ...).

For each handler, there are usually 3 layers: handler, service, repository:

• handler interacts directly with gRPC, REST or consumer using specific codes (cookies, ...) In case gRPC, there are frameworks outside handle for us so we can write business/logic codes here too. But remember, gRPC only.
• service is where we write business/logic codes, and only business/logic codes is written here.
• repository is where we write codes which interacts with database/cache like MySQL, Redis, ...
• models is where we put all request, response, data models.

Location:

• handler must exist inside grpc, http, consumer.
• service, models can exist directly inside of business if both grpc, http, consumer has same business/logic.
• repository should be placed directly inside of business.

If we have too many services, some of the logic will be overlapped.

For example, service A and service B both need to make POST call API to service C. If service A and service B both have libs to call service C to do that API, we need to move the libs to some common pkg libs. So in the future, service D which needs to call C will not need to copy libs to handle service C api but only need to import from common pkg libs.

Another bad practice is adapter service. No need to write a new service if what we need is just common pkg libs.

If I see someone using global var, I swear I will shoot them twice in the face.

Why?

• Can not write unit test.
• Is not thread safe.

For simple struct with 1 or 2 fields, no need to use functional options.

Example:

func main() {
	s := NewS(WithA(1), WithB("b"))
	fmt.Printf("%+v\n", s)
}

type S struct {
	fieldA int
	fieldB string
}

type OptionS func(s *S)

func WithA(a int) OptionS {
	return func(s *S) {
		s.fieldA = a
	}
}

func WithB(b string) OptionS {
	return func(s *S) {
		s.fieldB = b
	}
}

func NewS(opts ...OptionS) *S {
	s := &S{}
	for _, opt := range opts {
		opt(s)
	}
	return s
}

In above example, I construct s with WithA and WithB option. No need to pass direct field inside s.

If business logic involves calling too many APIs, but they are not depend on each other. We can fire them parallel :)

Personally, I prefer errgroup to WaitGroup (https://pkg.go.dev/sync#WaitGroup). Because I always need deal with error. Be super careful with egCtx, should use this instead of parent ctx inside eg.Go.

Example:

eg, egCtx := errgroup.WithContext(ctx)

eg.Go(func() error {
	// Do some thing
	return nil
})

eg.Go(func() error {
	// Do other thing
	return nil
})

if err := eg.Wait(); err != nil {
	// Handle error
}

Please don't use external libs for WorkerPool, I don't want to deal with dependency hell.

Example:

var bufPool = sync.Pool{
	New: func() any {
		return new(bytes.Buffer)
	},
}

func MarshalWithoutEscapeHTML(v any) ([]byte, error) {
	b, ok := bufPool.Get().(*bytes.Buffer)
	if !ok {
		return nil, ErrBufPoolNotBytesBuffer
	}

	b.Reset()
	defer bufPool.Put(b)

	encoder := json.NewEncoder(b)
	encoder.SetEscapeHTML(false)
	if err := encoder.Encode(v); err != nil {
		return nil, err
	}

	result := make([]byte, b.Len())
	copy(result, b.Bytes())
	return result, nil
}

Take value then return pointer, useful with database struct full of pointers:

// Ptr takes in non-pointer and returns a pointer
func Ptr[T any](v T) *T {
	return &v
}

Return zero value:

func Zero[T any]() T {
  var zero T
  return zero
}

Only need if you need something from vendor, to generate mock or something else.

To easily control version of build tools.

For example build.go:

//go:build tools
// +build tools

package main

import (
	_ "github.com/golang/protobuf/protoc-gen-go"
)

And then in Makefile:

build:
    go install github.com/golang/protobuf/protoc-gen-go

We always get the version of build tools in go.mod each time we install it. Future contributors will not cry anymore.

What is the point to pass many params (do-it, --abc, --xyz) when what we only need is start service?

In my case, service starts with only config, and config should be read from file or environment like The Twelve Factors guide.

Just don't.

Use protocolbuffers/protobuf-go, grpc/grpc-go for gRPC.

Write 1 for both gRPC, REST sounds good, but in the end, it is not worth it.

prototool is deprecated, and buf can generate, lint, format as good as prototool.

With c *gin.Context:

• Don't use c when passing context, use c.Request.Context() instead.
• Don't use c.Request.URL.Path, use c.FullPath() instead.

Remember to free resources after parse multipart form:

defer func() {
    if err := c.Request.MultipartForm.RemoveAll(); err != nil {
        fmt.Println(err)
    }
}()

It is fast!

• Don't overuse func (*Logger) With. Because if log line is too long, there is a possibility that we can lost it.
• Use MarshalLogObject when we need to hide some field of object when log (field is long or has sensitive value)
• Don't use Panic. Use Fatal for errors when start service to check dependencies. If you really need panic level, use DPanic.
• If doubt, use zap.Any.
• Use contextID or traceID in every log lines for easily debug.

Only init config in main or cmd layer. Do not use viper.Get... in business layer or inside business layer.

Why?

• Hard to mock and test
• Put all config in single place for easily tracking

Also, be careful if config value is empty. You should decide to continue or stop the service if there is empty config.

Each ORM libs has each different syntax. To learn and use those libs correctly is time consuming. So just stick to plain SQL. It is easier to debug when something is wrong.

Also please use prepared statement as much as possible. Idealy, we should init all prepared statement when we init database connection to cached it, not create it every time we need it.

But database/sql has its own limit. For example, it is hard to get primary key after insert/update. So may be you want to use ORM for those cases. I hear that go-gorm/gorm is good.

Be careful when use HGETALL. If key not found, empty data will be returned not nil error. See redis/go-redis/issues/1668

Use Pipelines for:

• HSET and EXPIRE in 1 command.
• Multiple GET in 1 command.

Prefer to use Pipelined instead of Pipeline. Inside Pipelined, please return redis.Cmder for each command.

Example:

func (c *client) HSetWithExpire(ctx context.Context, key string, values []any, expired time.Duration) error {
	cmds := make([]redis.Cmder, 2)

	if _, err := c.Pipelined(ctx, func(pipe redis.Pipeliner) error {
		cmds[0] = pipe.HSet(ctx, key, values...)

		if expired > 0 {
			cmds[1] = pipe.Expire(ctx, key, expired)
		}

		return nil
	}); err != nil {
		return err
	}

	for _, cmd := range cmds {
		if cmd == nil {
			continue
		}

		if err := cmd.Err(); err != nil {
			return err
		}
	}

	return nil
}

Remember to config:

• ReadTimeout, WriteTimeout

Remember to config:

• SetConnMaxLifetime
• SetMaxOpenConns
• SetMaxIdleConns
• ParseTime to true.
• Loc to time.UTC.
• CheckConnLiveness to true.
• ReadTimeout, WriteTimeout

Remember to config:

• Write-Ahead Logging: PRAGMA journal_mode=WAL
• Disable connections pool with SetMaxOpenConns sets to 1

Don't use mattn/go-sqlite3, it's required CGO_ENABLED.

Use sarama.V1_0_0_0, because IBM decide to upgrade default version.

Don't use confluentinc/confluent-kafka-go, it's required CGO_ENABLED.

It is easy to write a suite test, thanks to testify. Also, for mocking, there are many options out there. Pick 1 then sleep peacefully.

The first is easy to use but not powerful as the later. If you want to make sure mock func is called with correct times, use the later.

Example with matryer/moq:

// Only gen mock if source code file is newer than mock file
// https://jonwillia.ms/2019/12/22/conditional-gomock-mockgen
//go:generate sh -c "test service_mock_generated.go -nt $GOFILE && exit 0; moq -rm -out service_mock_generated.go . Service"

Don't cast proto enum:

// Bad
a := cast.ToInt32(servicev1.ReasonCode_ABC)

// Good
a := int32(servicev1.ReasonCode_ABC)

type Drink int

const (
	Beer Drink = iota
	Water
	OrangeJuice
)

go install golang.org/x/tools/cmd/stringer@latest

# Run inside directory which contains Drink
stringer -type=Drink

rate if you want rate limiter locally in your single instance of service. redis_rate if you want rate limiter distributed across all your instances of service.

gofumpt provides more rules when format Go codes.

No need to say more. Lint or get the f out!

If you get fieldalignment error, use fieldalignment to fix them.

# Install
go install golang.org/x/tools/go/analysis/passes/fieldalignment/cmd/fieldalignment@latest

# Fix
fieldalignment -fix ./internal/business/*.go

Change import:

gofmt -w -r '"github.com/Sirupsen/logrus" -> "github.com/sirupsen/logrus"' *.go

Cleanup if storage is full:

go clean -cache -testcache -modcache -fuzzcache -x

• Functional options for friendly APIs
• Designing Go Libraries: The Talk: The Article
• Uber Go Style Guide
• Google Go Style
• Three bugs in the Go MySQL Driver
• Fixing Memory Exhaustion Bugs in My Golang Web App
• Prevent Logging Secrets in Go by Using Custom Types
• Speed Up GoMock with Conditional Generation
• Making SQLite faster in Go
• Advanced Go Concurrency
• Go generic: non-ptr to ptr
• Crimes with Go Generics