diff --git a/cluster/core/core.go b/cluster/core/core.go
index d1f792c..dbf3627 100644
--- a/cluster/core/core.go
+++ b/cluster/core/core.go
@@ -113,6 +113,29 @@ func NewCluster(cfg *Config) (*Cluster, error) {
 	if err != nil {
 		return nil, err
 	}
+
+	cluster := &Cluster{
+		raftNode:        raftNode,
+		db:              db,
+		connections:     connections,
+		config:          cfg,
+		rebalanceManger: newRebalanceManager(),
+		slotsManager:    newSlotsManager(),
+		transactions:    newTransactionManager(),
+		replicaManager:  newReplicaManager(),
+		closeChan:       make(chan struct{}),
+	}
+	cluster.pickNodeImpl = func(slotID uint32) string {
+		return defaultPickNodeImpl(cluster, slotID)
+	}
+	cluster.getSlotImpl = func(key string) uint32 {
+		return defaultGetSlotImpl(cluster, key)
+	}
+	cluster.injectInsertCallback()
+	cluster.injectDeleteCallback()
+	cluster.registerOnFailover()
+
+	// setup
 	hasState, err := raftNode.HasExistingState()
 	if err != nil {
 		return nil, err
@@ -140,34 +163,23 @@ func NewCluster(cfg *Config) (*Cluster, error) {
 				return nil, err
 			}
 		}
+	} else {
+		masterAddr := cluster.raftNode.FSM.GetMaster(cluster.SelfID())
+		if masterAddr != "" {
+			err := cluster.SlaveOf(masterAddr)
+			if err != nil {
+				panic(err)
+			}
+		}
 	}
-	cluster := &Cluster{
-		raftNode:        raftNode,
-		db:              db,
-		connections:     connections,
-		config:          cfg,
-		rebalanceManger: newRebalanceManager(),
-		slotsManager:    newSlotsManager(),
-		transactions:    newTransactionManager(),
-		replicaManager:  newReplicaManager(),
-		closeChan:       make(chan struct{}),
-	}
-	cluster.pickNodeImpl = func(slotID uint32) string {
-		return defaultPickNodeImpl(cluster, slotID)
-	}
-	cluster.getSlotImpl = func(key string) uint32 {
-		return defaultGetSlotImpl(cluster, key)
-	}
-	cluster.injectInsertCallback()
-	cluster.injectDeleteCallback()
-	cluster.registerOnFailover()
+	
 	go cluster.clusterCron()
 	return cluster, nil
 }
 
 // AfterClientClose does some clean after client close connection
 func (cluster *Cluster) AfterClientClose(c redis.Connection) {
-
+	cluster.db.AfterClientClose(c)
 }
 
 func (cluster *Cluster) Close() {
diff --git a/cluster/core/replica_manager.go b/cluster/core/replica_manager.go
index bcf428b..1ce5cec 100644
--- a/cluster/core/replica_manager.go
+++ b/cluster/core/replica_manager.go
@@ -13,6 +13,14 @@ import (
 	"github.com/hdt3213/godis/redis/protocol"
 )
 
+/*
+1. The master and slave are both added to the raft group, and failover does not involve changes of raft members. 
+2. Timer job `doFailoverCheck` finds timeout masters, then calls `triggerFailover`
+3. Raft leader sends `slaveof no one` to new master
+4. Raft proposes `EventFinishFailover` to change route. 
+Other slaves and old master will get this message from raft, and become slave of new master.(see cluster.registerOnFailover)
+*/
+
 const heartbeatCommand = "cluster.heartbeat"
 
 func init() {
@@ -93,19 +101,7 @@ func (cluster *Cluster) doFailoverCheck() {
 func (cluster *Cluster) triggerFailover(failed *raft.MasterSlave) error {
 	newMaster := failed.Slaves[0]
 	id := utils.RandString(20)
-	// propose change
-	_, err := cluster.raftNode.Propose(&raft.LogEntry{
-		Event: raft.EventStartFailover,
-		FailoverTask: &raft.FailoverTask{
-			ID:          id,
-			OldMasterId: failed.MasterId,
-			NewMasterId: newMaster,
-		},
-	})
-	if err != nil {
-		return err
-	}
-	logger.Infof("proposed start failover id=%s, oldMaster=%s, newMaster=%s", id, failed.MasterId, newMaster)
+	logger.Infof("start failover id=%s, oldMaster=%s, newMaster=%s", id, failed.MasterId, newMaster)
 	// send slave of to new master
 	conn, err := cluster.connections.BorrowPeerClient(newMaster)
 	if err != nil {
diff --git a/cluster/core/utils.go b/cluster/core/utils.go
index 9b9816e..dc74006 100644
--- a/cluster/core/utils.go
+++ b/cluster/core/utils.go
@@ -1,10 +1,14 @@
 package core
 
 import (
+	"errors"
 	"hash/crc32"
+	"net"
 	"strings"
 
 	"github.com/hdt3213/godis/interface/redis"
+	"github.com/hdt3213/godis/lib/utils"
+	"github.com/hdt3213/godis/redis/connection"
 	"github.com/hdt3213/godis/redis/protocol"
 )
 
@@ -44,6 +48,19 @@ func (cluster *Cluster) LocalExecWithinLock(c redis.Connection, cmdLine [][]byte
 	return cluster.db.ExecWithLock(c, cmdLine)
 }
 
+func (cluster *Cluster) SlaveOf(master string) error {
+	host, port, err := net.SplitHostPort(master)
+	if err != nil {
+		return errors.New("invalid master address")
+	}
+	c := connection.NewFakeConn()
+	reply := cluster.db.Exec(c, utils.ToCmdLine("slaveof", host, port))
+	if err := protocol.Try2ErrorReply(reply); err != nil {
+		return err
+	}
+	return nil
+}
+
 // GetPartitionKey extract hashtag
 func GetPartitionKey(key string) string {
 	beg := strings.Index(key, "{")
diff --git a/cluster/raft/fsm.go b/cluster/raft/fsm.go
index f0f61f5..180b0bf 100644
--- a/cluster/raft/fsm.go
+++ b/cluster/raft/fsm.go
@@ -67,7 +67,6 @@ const (
 	EventStartMigrate = iota + 1
 	EventFinishMigrate
 	EventSeedStart
-	EventStartFailover
 	EventFinishFailover
 	EventJoin
 )
@@ -109,9 +108,6 @@ func (fsm *FSM) Apply(log *raft.Log) interface{} {
 		}
 		fsm.Node2Slot[entry.InitTask.Leader] = slots
 		fsm.addNode(entry.InitTask.Leader, "")
-	} else if entry.Event == EventStartFailover {
-		task := entry.FailoverTask
-		fsm.Failovers[task.ID] = task
 	} else if entry.Event == EventFinishFailover {
 		task := entry.FailoverTask
 		// change route
diff --git a/cluster/raft/fsm_utils.go b/cluster/raft/fsm_utils.go
index 54514f8..6bf6c93 100644
--- a/cluster/raft/fsm_utils.go
+++ b/cluster/raft/fsm_utils.go
@@ -79,8 +79,9 @@ func (fsm *FSM) failover(oldMasterId, newMasterId string) {
 	}
 }
 
-// getMaster returns "" if id points to a master node
-func (fsm *FSM) getMaster(id string) string {
+// GetMaster returns the master's redis service addres 
+// Returns empty string ("") if id points to a master node
+func (fsm *FSM) GetMaster(id string) string {
 	master := ""
 	fsm.WithReadLock(func(fsm *FSM) {
 		master = fsm.SlaveMasters[id]
@@ -111,4 +112,4 @@ func (fsm *FSM) addNode(id, masterId string) error {
 		fsm.SlaveMasters[id] = masterId
 	}
 	return nil
-}
\ No newline at end of file
+}
diff --git a/cluster/raft/raft.go b/cluster/raft/raft.go
index c6bc22a..63e0e9c 100644
--- a/cluster/raft/raft.go
+++ b/cluster/raft/raft.go
@@ -191,6 +191,7 @@ func (node *Node) Propose(event *LogEntry) (uint64, error) {
 	return future.Index(), nil
 }
 
+// setupWatch will be called after any changes of FSM within lock.
 func (node *Node) setupWatch() {
 	node.watcher.watch = func(f *FSM) {
 		newMaster := f.SlaveMasters[node.Self()]
@@ -206,6 +207,6 @@ func (node *Node) setupWatch() {
 // SetOnFailover sets onFailover callback
 // After a failover, onFailover will receive the new master
 func (node *Node) SetOnFailover(fn func(newMaster string)) {
-	node.watcher.currentMaster = node.FSM.getMaster(node.Self())
+	node.watcher.currentMaster = node.FSM.GetMaster(node.Self())
 	node.watcher.onFailover = fn
 }