diff --git a/messaging/fanout.go b/messaging/fanout.go
new file mode 100644
index 0000000..e73240b
--- /dev/null
+++ b/messaging/fanout.go
@@ -0,0 +1,144 @@
+package messaging
+
+import (
+	"context"
+	"io"
+	"sync"
+	"sync/atomic"
+
+	"bitbucket.org/sakariai/sakari/log"
+	"bitbucket.org/sakariai/sakari/log/field"
+)
+
+const (
+	MaxWorkers   = 32
+	MaxQueueSize = 128
+)
+
+var (
+	running uint32 = 0
+)
+
+type Pipeline struct {
+	workers []*worker
+	chain   chan interface{}
+}
+
+func (p *Pipeline) Start() {
+	distributeToChannels := func(ch chan interface{}, cs []*worker) {
+		writer := cs[0] //first worker must stream as default
+		for {
+			for _, c := range cs {
+				expectationWorkers := uint32(len(ch)/(MaxQueueSize/MaxWorkers)) + 1
+				select {
+				case val := <-ch:
+					runningWorker := atomic.LoadUint32(&running)
+					if c.index <= runningWorker || c.index <= expectationWorkers {
+						writer = c
+					}
+					if c.debug {
+						log.Info("Worker receiving", field.Any("index", writer.index), field.Any("running", runningWorker), field.Any("no# workers", expectationWorkers))
+					}
+					go writer.stream(val)
+				}
+			}
+		}
+	}
+
+	go distributeToChannels(p.chain, p.workers)
+}
+
+func (p *Pipeline) Dispatch(msg interface{}) {
+	p.chain <- msg
+}
+
+type DispatcherBuilder func() Dispatcher
+
+func NewPipeline(d DispatcherBuilder, ch chan interface{}, idle uint32, debug bool) *Pipeline {
+	wk := make([]*worker, 0, MaxWorkers)
+	for i := 0; i < MaxWorkers; i++ {
+		wk = append(wk,
+			&worker{
+				index:      uint32(i + 1),
+				chain:      make(chan interface{}, MaxQueueSize),
+				mutex:      new(sync.Mutex),
+				debug:      debug,
+				idle:       idle,
+				Dispatcher: d(),
+			})
+	}
+	return &Pipeline{workers: wk, chain: ch}
+}
+
+type Dispatcher interface {
+	Before(context.Context) error
+	After() error
+	Process(interface{}) error
+}
+
+type worker struct {
+	index   uint32
+	mutex   *sync.Mutex
+	running bool
+	chain   chan interface{}
+	debug   bool
+	idle    uint32
+	Dispatcher
+}
+
+func (c *worker) stream(val interface{}) {
+	c.chain <- val
+	if !c.running {
+		c.mutex.Lock()
+		c.running = true
+		atomic.AddUint32(&running, 1)
+		defer atomic.AddUint32(&running, ^uint32(1-1))
+		ctx, cancel := context.WithCancel(context.Background())
+		err := c.Before(ctx)
+
+		if err != nil {
+			log.Error("can not start worker", field.Error(err))
+		}
+		defer func(w *worker, cancel context.CancelFunc) {
+			if w.debug {
+				log.Info("Worker leaving", field.Any("index", w.index), field.Any("idle", w.idle))
+			}
+			err := c.After()
+			if err != nil {
+				log.Error("can not finish track issue", field.Error(err))
+			}
+			cancel()
+			w.mutex.Unlock()
+			w.running = false
+		}(c, cancel)
+		var idle uint32 = 0
+		for {
+			select {
+			case msg := <-c.chain:
+				atomic.StoreUint32(&idle, 0)
+				if msg != nil {
+					err := c.Process(msg)
+					if err != nil {
+						log.Error("can not process message",
+							field.Any("msg", &msg),
+							field.Error(err),
+						)
+					}
+					if err == io.EOF {
+						return
+					}
+				}
+			default:
+				atomic.AddUint32(&idle, 1)
+				if i := atomic.LoadUint32(&idle); i > 0 {
+					if i > c.idle {
+						return
+					}
+					if c.debug {
+						log.Info("Idle", field.Any("worker index", c.index), field.Any("idle", idle))
+					}
+				}
+			}
+		}
+	}
+}
diff --git a/messaging/fanout.md b/messaging/fanout.md
new file mode 100644
index 0000000..06541d2
--- /dev/null
+++ b/messaging/fanout.md
@@ -0,0 +1,205 @@
+## Implementation 
+We can activate worker base on traffic of parent channel
+
+```go
+package concurrency
+
+import (
+	"context"
+	"io"
+	"sync"
+	"sync/atomic"
+
+	"bitbucket.org/sakariai/sakari/log"
+	"bitbucket.org/sakariai/sakari/log/field"
+)
+
+const (
+	MaxWorkers   = 32
+	MaxQueueSize = 128
+)
+
+var (
+	running uint32 = 0
+)
+
+type Pipeline struct {
+	workers []*worker
+	chain   chan interface{}
+}
+
+func (p *Pipeline) Start() {
+	distributeToChannels := func(ch chan interface{}, cs []*worker) {
+		writer := cs[0] //first worker must stream as default
+		for {
+			for _, c := range cs {
+				expectationWorkers := uint32(len(ch)/(MaxQueueSize/MaxWorkers)) + 1
+				select {
+				case val := <-ch:
+					runningWorker := atomic.LoadUint32(&running)
+					if c.index <= runningWorker || c.index <= expectationWorkers {
+						writer = c
+					}
+					if c.debug {
+						log.Info("Worker receiving", field.Any("index", writer.index), field.Any("running", runningWorker), field.Any("no# workers", expectationWorkers))
+					}
+					go writer.stream(val)
+				}
+			}
+		}
+	}
+
+	go distributeToChannels(p.chain, p.workers)
+}
+
+func (p *Pipeline) Dispatch(msg interface{}) {
+	p.chain <- msg
+}
+
+type DispatcherBuilder func() Dispatcher
+
+func NewPipeline(d DispatcherBuilder, ch chan interface{}, idle uint32, debug bool) *Pipeline {
+	wk := make([]*worker, 0, MaxWorkers)
+	for i := 0; i < MaxWorkers; i++ {
+		wk = append(wk,
+			&worker{
+				index:      uint32(i + 1),
+				chain:      make(chan interface{}, MaxQueueSize),
+				mutex:      new(sync.Mutex),
+				debug:      debug,
+				idle:       idle,
+				Dispatcher: d(),
+			})
+	}
+	return &Pipeline{workers: wk, chain: ch}
+}
+
+type Dispatcher interface {
+	Before(context.Context) error
+	After() error
+	Process(interface{}) error
+}
+
+type worker struct {
+	index   uint32
+	mutex   *sync.Mutex
+	running bool
+	chain   chan interface{}
+	debug   bool
+	idle    uint32
+	Dispatcher
+}
+
+func (c *worker) stream(val interface{}) {
+	c.chain <- val
+	if !c.running {
+		c.mutex.Lock()
+		c.running = true
+		atomic.AddUint32(&running, 1)
+		defer atomic.AddUint32(&running, ^uint32(1-1))
+		ctx, cancel := context.WithCancel(context.Background())
+		err := c.Before(ctx)
+
+		if err != nil {
+			log.Error("can not start worker", field.Error(err))
+		}
+		defer func(w *worker, cancel context.CancelFunc) {
+			if w.debug {
+				log.Info("Worker leaving", field.Any("index", w.index), field.Any("idle", w.idle))
+			}
+			err := c.After()
+			if err != nil {
+				log.Error("can not finish track issue", field.Error(err))
+			}
+			cancel()
+			w.mutex.Unlock()
+			w.running = false
+		}(c, cancel)
+		var idle uint32 = 0
+		for {
+			select {
+			case msg := <-c.chain:
+				atomic.StoreUint32(&idle, 0)
+				if msg != nil {
+					err := c.Process(msg)
+					if err != nil {
+						log.Error("can not process message",
+							field.Any("msg", &msg),
+							field.Error(err),
+						)
+					}
+					if err == io.EOF {
+						return
+					}
+				}
+			default:
+				atomic.AddUint32(&idle, 1)
+				if i := atomic.LoadUint32(&idle); i > 0 {
+					if i > c.idle {
+						return
+					}
+					if c.debug {
+						log.Info("Idle", field.Any("worker index", c.index), field.Any("idle", idle))
+					}
+				}
+			}
+		}
+	}
+}
+
+```
+
+## Usage
+
+```go
+import concurrency
+
+type taggingDispatcher struct {
+	Address string
+	stream  proto.Havilah_StreamMetricClient
+	conn    *grpc.ClientConn
+}
+
+func (d *taggingDispatcher) Before(ctx context.Context) error {
+	conn, err := grpc.Dial(d.Address, grpc.WithInsecure())
+	if err != nil {
+		return err
+	}
+	d.conn = conn
+	client := proto.NewHavilahClient(conn)
+
+	stream, err := client.StreamMetric(ctx)
+	if err != nil {
+		return err
+	}
+	d.stream = stream
+	return nil
+}
+
+func (d *taggingDispatcher) After() error {
+	_, err := d.stream.CloseAndRecv()
+
+	e := d.conn.Close()
+	if e != nil {
+		log.Error("close havilah connection error", field.Error(e))
+	}
+	return err
+}
+
+func (d *taggingDispatcher) Process(msg interface{}) error {
+	return d.stream.Send(msg.(*proto.Tagging))
+}
+
+
+tagging := &Tagging{
+    topic: topic,
+    pipeline: concurrency.NewPipeline(func() concurrency.Dispatcher {
+        return &taggingDispatcher{Address: address}
+    }, ch, idle, debug),
+}
+tagging.pipeline.Start()
+
+func main(){
+	tagging.pipeline.Dispatch(youStruct{})
+}
+```