YALA - Yet Another Logging Abstraction for Go
Tiny structured logging abstraction or facade for various logging libraries, allowing the end user to plug in the desired logging library in main.go.
Supported logging libraries (via adapters)
logrus, zap, zerolog, glog, log15, standard log and console
When to use?
- If you are a package/module/library author
- And you want to participate in the end user logging system (log messages using the logger provided by the end user)
- You don't want to add dependency to any specific logging library to your code
- You don't want to manually inject logger to every possible place where you want to log something (such as function, struct etc.)
- If you need a nice and elegant API with a bunch of useful functions, but at the same time you don't want your end users to spend hours on writing their own logging adapter.
Installation
# Add yala to your Go module:
go get github.com/elgopher/yala
Please note that at least Go 1.17 is required.
How to use
Choose logger - global or local?
Global logger can be accessed from everywhere in your library and can be reconfigured anytime. Local logger is a logger initialized only once and used locally, for example inside the function.
Use global logger
package lib // this is your package, part of module/library etc.
import (
"context"
"errors"
"github.com/elgopher/yala/logger"
)
// define global logger, no need to initialize it (by default nothing is logged)
var log logger.Global
// Provide a public function for setting adapter. It will be called in main.go
func SetLoggerAdapter(adapter logger.Adapter) {
log.SetAdapter(adapter)
}
func Function(ctx context.Context) {
log.Debug(ctx, "Debug message")
log.With("field_name", "value").Info(ctx, "Message with field")
log.WithError(errors.New("some")).Error(ctx, "Message with error")
}
Specify adapter - a real logger implementation.
package main
import (
"context"
"github.com/elgopher/yala/adapter/console"
"lib"
)
// End user decides what library to plug in.
func main() {
adapter := console.StdoutAdapter() // will print messages to console
lib.SetLoggerAdapter(adapter)
ctx := context.Background()
lib.Function(ctx)
}
Why context.Context is a parameter?
context.Context can very useful in transiting request-scoped tags or even entire logger. A logger.Adapter implementation might use them making possible to log messages instrumented with tags. Thanks to that your library can trully participate in the incoming request.
Use local logger
Logging is a special kind of dependency. It is used all over the place. Adding it as an explicit dependency to every function, struct etc. can be cumbersome. Still though, you have an option to use local logger by injecting logger.Adapter into your library:
// your library code:
func NewLibrary(adapter logger.Adapter) YourLib {
// create a new local logger which provides similar API to the global logger
localLogger := logger.Local{Adapter: adapter}
return YourLib{localLogger: localLogger}
}
type YourLib struct {
localLogger logger.Local
}
func (l YourLib) Method(ctx context.Context) {
l.localLogger.Debug(ctx, "message from local logger")
}
// end user code
adapter := console.StdoutAdapter()
lib := NewLibrary(adapter)
How to use existing adapters
Writing your own adapter
Just implement logger.Adapter interface:
type MyAdapter struct{}
func (MyAdapter) Log(context.Context, logger.Entry) {
// here you can do whatever you want with the log entry
}
Difference between Logger and Adapter
- Logger is used by package/module/library author
- Adapter is an interface to be implemented by adapters. They use real logging libraries under the hood.
- So, why two abstractions? Simply because the smaller the Adapter interface, the easier it is to implement it. On the other hand, from library perspective, more methods means API which is easier to use.
- Here is the architecture from the package perspective:
More examples
Advanced recipes
- Filter out messages starting with given prefix
- Filter messages by level
- Add field to each message taken from context.Context
- Zap logger passed over context.Context
Why just don't create my own abstraction instead of using yala?
Yes, you can also create your own. Very often it is just an interface with a single method, like this:
type ImaginaryLogger interface {
Log(context.Context, Entry)
}
But there are limitations for such solution:
- such interface alone is not very easy to use in your package/module/library. You just have to write way too much boilerplate code.
- someone who is using your package is supposed to write implementation of this interface (or you can provide prebuilt implementation for various logging libraries). In both cases this cost time and effort.
- it is not obvious how logging API should look like. Someone would argue that is better to have a much more complicated interface like this:
type AnotherImaginaryLogger interface {
With(field string, value interface{}) AnotherImaginaryLogger
WithError(err error) AnotherImaginaryLogger
Info(context.Context, string)
Debug(context.Context, string)
Warn(context.Context, string)
Error(context.Context, string)
}
Unfortunately such interface is much harder to implement, than interface with a single method.
But yala is just another API. Why it is unique?
- yala is designed for the ease of use. And by that I mean ease of use for everyone - developer logging messages, developer writing adapter and end user configuring the adapter
- yala is using
context.Contextin each method call, making possible to use sophisticated request-scoped logging
YALA limitations
- even though your package will be independent of any specific logging implementation, you still have to import
github.com/elgopher/yala/logger. This package is relatively small though, compared to real logging libraries (about ~200 lines of production code) and it does not import any external libraries. - yala is not optimized for extreme performance, because this would hurt the developer experience and readability of the created code. Any intermediary API ads overhead - global synchronized variables, wrapper code and even polymorphism slow down the execution a bit. The overhead varies, but it is usually a matter of tens of nanoseconds per call.
