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_1
+| business
| | http
| | | handler.go
| | | service.go
-| | | repository.go
| | | models.go
| | grpc
| | | handler.go
+| | | models.go
+| | consumer
+| | | handler.go
| | | service.go
-| | | repository.go
| | | models.go
| | service.go
| | repository.go
| | models.go
-| business_2
-| | grpc
-| | | handler.go
-| | | service.go
-| | | repository.go
-| | | models.go
-All business codes are inside internal.
Each business has a different directory (business_1, business_2).
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).Inside each handler, there are usually 3 layers: handler, service, repository:
handler interacts directly with gRPC or REST using specific codes (cookies,...)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, ...handler must exist inside grpc, http.
But service, repository, models can exist directly inside business if both grpc, http has same business/logic.
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.
For simple struct with 1 or 2 fields, no need to use functional options.
func main() {
+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, ... And we should place it directly inside of business.models is where we put all request, response, data models.handler must exist inside grpc, http.
But service, models can exist directly inside of business if both grpc, http has same business/logic.
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.
For simple struct with 1 or 2 fields, no need to use functional options.
func main() {
s := NewS(WithA(1), WithB("b"))
fmt.Printf("%+v\n", s)
}
@@ -51,4 +46,9 @@ func NewS(opts ...OptionS) *S {
}
return s
}
-In above example, I construct s with WithA and WithB option.
No need to pass direct field inside s.
vendorOnly need if you need something from vendor, to generate mock or something else.
What is the point to pass many params (--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.
Don't use gin.Context when pass context from handler layer to service layer, use gin.Context.Request.Context() instead.
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 has long or 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.
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.
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.
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.
go fmt, goimports with mvdan/gofumpt.gofumpt provides more rules when format Go codes.
No need to say more.
Lint or get the f out!
In above example, I construct s with WithA and WithB option.
No need to pass direct field inside s.
vendorOnly need if you need something from vendor, to generate mock or something else.
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.
Don't use gin.Context when pass context from handler layer to service layer, use gin.Context.Request.Context() instead.
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.
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.
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.
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.
go fmt, goimports with mvdan/gofumpt.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@latest
+
+# Fix
+fieldalignment -fix ./internal/business/*.go
+