kamilsk / tracer

🪡 Dead simple, lightweight tracing.

🧶 tracer

Dead simple, lightweight tracing.

💡 Idea

The tracer provides API to trace execution flow.

func Do(ctx context.Context) {
	defer tracer.Fetch(ctx).Start().Stop()

	// do some heavy job
}

A full description of the idea is available here.

🏆 Motivation

At Avito, we use the Jaeger - a distributed tracing platform. It is handy in most cases, but at production, we also use sampling. So, what is a problem, you say?

I had 0.02% requests with a write: broken pipe error and it was difficult to find the appropriate one in the Sentry which also has trace related to it in the Jaeger.

For that reason, I wrote the simple solution to handle this specific case and found the bottleneck in our code quickly.

🤼‍♂️ How to

import (
	"context"
	"io"
	"net/http"
	"time"

	"github.com/kamilsk/tracer"
)

func InjectTracer(handler http.Handler) http.Handler {
	return http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		req = req.WithContext(tracer.Inject(req.Context(), make([]*tracer.Call, 0, 10)))
		handler.ServeHTTP(rw, req)
	})
}

func Handle(rw http.ResponseWriter, req *http.Request) {
	ctx, cancel := context.WithTimeout(req.Context(), time.Second)
	defer cancel()

	call := tracer.Fetch(req.Context()).Start(req.Header.Get("X-Request-Id"))
	defer call.Stop()

	...

	call.Checkpoint("serialize")
	data, err := FetchData(ctx, req.Body)
	if err != nil {
		http.Error(rw, err.Error(), http.StatusInternalServerError)
		return
	}

	call.Checkpoint("store")
	if err := StoreIntoDatabase(ctx, data); err != nil {
		http.Error(rw,
			http.StatusText(http.StatusInternalServerError),
			http.StatusInternalServerError)
		return
	}
	rw.WriteHeader(http.StatusOK)
}

func FetchData(ctx context.Context, r io.Reader) (Data, error) {
	defer tracer.Fetch(ctx).Start().Stop()

	// fetch a data into a struct
}

func StoreIntoDatabase(ctx context.Context, data Data) error {
	defer tracer.Fetch(ctx).Start().Stop()

	// store the data into a database
}

Output:

allocates at call stack: 0, detailed call stack:
	call Handle [ca7a87c4-58d0-4fdf-857c-ef49fc3bf271]: 14.038083ms, allocates: 2
		checkpoint [serialize]: 1.163587ms
		checkpoint [store]: 2.436265ms
	call FetchData: 1.192829ms, allocates: 0
	call StoreIntoDatabase: 10.428663ms, allocates: 0

🧩 Integration

The library uses SemVer for versioning, and it is not BC-safe through major releases. You can use go modules to manage its version.

$ go get github.com/kamilsk/tracer@latest

---

made with ❤️ for everyone

tracer's Issues

start time for Inject operation

mark start operation on timeline

http calls dump

• record requests
• dump them to *.http for IDE
• build call graph

improve Stop()

• add it to tracer
• add it to call

research httptrace and others

• https://blog.golang.org/http-tracing
• ...

up go version

go 1.13

make it thread safe

now it's not

create mirror at bitbucket

isolate strings.Builder for printer

to fix builds on go 1.7, 1.8, 1.9

https://travis-ci.org/kamilsk/tracer/jobs/575204658

Checkpoint as key value.

Can we support something like this?

Checkpoint("key", "value")

bucket to store debug info

call.Store(Request{
  Type: HTTP
  ...
})

call.Store(Request{
  Type: SQL
  ...
})

example with sla

func Middleware(handler http.Handler) http.Handler {
  return http.HandlerFunc(func (...) {
    ctx := context.WithValue(req.Context(), key, tracer)
    http.ServeHTTP(rw, req.WithContex(ctx))
    if ctx.Err() != nil {
      // fetch tracer, write trace into Sentry
    }
  })
}

add possibility to preallocate checkpoints

call := tracer.Fetch(ctx).With(make([]Checkpoint, 0, 10)).Start()
defer call.Stop()
...
call.Checkpoint()
...
call.Checkpoint("id", "labelX", "labelY")
...

add depth counter

add possibility to build a graph

• id for each node
• parent id for each child node
• blocker is goroutines support