(etcd)raft算法的学习笔记 (二)

2 篇文章 0 订阅
1 篇文章 0 订阅

Raft 一致性算法

section 5几个子章节结合etcd源码讨论
一、section 5.1

  • 服务器始终处于三种状态:leader、follower、candidate(etcd新增了一个learner,在后面补充)
  • follower: 该角色处于被动位置从不会主动发起请求,仅仅是简单的回应leader和candidate的请求
  • leader:该角色处理所有客户端发来的请求。如果follower不小心接收到了客户端请求,该请求将会转发给leader
  • candidate:该角色是follower到leader的过度状态
  • RAFT将时间划分为任意长度的任期,任期为连续的整数,每一任期从选举开始
  • 不同的servers能观察到多次任期的转换。任期在raft中充当逻辑时钟的作用,这可以帮助集群探查一些过期的信息例如旧的leader等。
  • 如果一个server的任期号比其他人的小,那么他会更新自己到较新的编号。如果此时处于leader或candidate状态,将会立即变味follower
//来自于 etcd/raft/raft_paper_test.go
func TestFollowerUpdateTermFromMessage(t *testing.T) {
    //对于follower更新到较新的编号
	testUpdateTermFromMessage(t, StateFollower)
}
raft2020/06/03 10:43:33 raft.go:1521: INFO: 1 switched to configuration voters=(1 2 3)
raft2020/06/03 10:43:33 raft.go:699: INFO: 1 became follower at term 0
raft2020/06/03 10:43:33 raft.go:383: INFO: newRaft 1 [peers: [1,2,3], term: 0, commit: 0, applied: 0, lastindex: 0, lastterm: 0]
raft2020/06/03 10:43:33 raft.go:699: INFO: 1 became follower at term 1
raft2020/06/03 10:43:33 raft.go:858: INFO: 1 [term: 1] received a MsgApp message with higher term from 0 [term: 2]
raft2020/06/03 10:43:33 raft.go:699: INFO: 1 became follower at term 2
func TestCandidateUpdateTermFromMessage(t *testing.T) {
    //对于candidate更新到较新的编号,最后状态会变为follower
	testUpdateTermFromMessage(t, StateCandidate)
}
raft2020/06/03 10:43:56 raft.go:1521: INFO: 1 switched to configuration voters=(1 2 3)
raft2020/06/03 10:43:56 raft.go:699: INFO: 1 became follower at term 0
raft2020/06/03 10:43:56 raft.go:383: INFO: newRaft 1 [peers: [1,2,3], term: 0, commit: 0, applied: 0, lastindex: 0, lastterm: 0]
raft2020/06/03 10:43:56 raft.go:712: INFO: 1 became candidate at term 1
raft2020/06/03 10:43:56 raft.go:858: INFO: 1 [term: 1] received a MsgApp message with higher term from 0 [term: 2]
raft2020/06/03 10:43:56 raft.go:699: INFO: 1 became follower at term 2
func TestLeaderUpdateTermFromMessage(t *testing.T) {
    //对于leader更新到较新的编号,最后状态会变为follower
	testUpdateTermFromMessage(t, StateLeader)
}
raft2020/06/03 10:44:27 raft.go:1521: INFO: 1 switched to configuration voters=(1 2 3)
raft2020/06/03 10:44:27 raft.go:699: INFO: 1 became follower at term 0
raft2020/06/03 10:44:27 raft.go:383: INFO: newRaft 1 [peers: [1,2,3], term: 0, commit: 0, applied: 0, lastindex: 0, lastterm: 0]
raft2020/06/03 10:44:27 raft.go:712: INFO: 1 became candidate at term 1
raft2020/06/03 10:44:27 raft.go:764: INFO: 1 became leader at term 1
raft2020/06/03 10:44:27 raft.go:858: INFO: 1 [term: 1] received a MsgApp message with higher term from 0 [term: 2]
raft2020/06/03 10:44:27 raft.go:699: INFO: 1 became follower at term 2
// testUpdateTermFromMessage tests that if one server’s current term is
// smaller than the other’s, then it updates its current term to the larger
// value. If a candidate or leader discovers that its term is out of date,
// it immediately reverts to follower state.
// Reference: section 5.1
func testUpdateTermFromMessage(t *testing.T, state StateType) {
	r := newTestRaft(1, []uint64{1, 2, 3}, 10, 1, NewMemoryStorage())   //初始化一个test raft,peers包括1、2、3
	switch state {
	case StateFollower:
		r.becomeFollower(1, 2) //1是给定的任期,2是leader的id
	case StateCandidate:
		r.becomeCandidate() //将初始化的raft状态变为candidate
	case StateLeader:
		r.becomeCandidate() //将初始化的raft状态由follower变为candidate,若follower直接转变为leader会出错
		r.becomeLeader() //candidate状态变为follower
	}

	r.Step(pb.Message{Type: pb.MsgApp, Term: 2})  //发送一个

	if r.Term != 2 {
		t.Errorf("term = %d, want %d", r.Term, 2)
	}
	if r.state != StateFollower {
		t.Errorf("state = %v, want %v", r.state, StateFollower)
	}
}
  • 如果一个server发现一个请求的任期号是旧的,这个请求将会给拒绝
// TestRejectStaleTermMessage tests that if a server receives a request with
// a stale term number, it rejects the request.
// Our implementation ignores the request instead.
// Reference: section 5.1
func TestRejectStaleTermMessage(t *testing.T) {
	called := false
	fakeStep := func(r *raft, m pb.Message) error {
		called = true
		return nil
	}
	r := newTestRaft(1, []uint64{1, 2, 3}, 10, 1, NewMemoryStorage())  //创建一个新的server
	//fmt.Printf("%+v,\n", r)
	r.step = fakeStep
	r.loadState(pb.HardState{Term: 2}) //将r的任期置为2
	//fmt.Println(r.Term)

	r.Step(pb.Message{Type: pb.MsgApp, Term: r.Term - 1}) //发送一个旧任期的请求给r

	if called {
		t.Errorf("stepFunc called = %v, want %v", called, false)
	}
}

=== RUN   TestRejectStaleTermMessage
raft2020/06/03 19:35:25 raft.go:1521: INFO: 1 switched to configuration voters=(1 2 3)
raft2020/06/03 19:35:25 raft.go:699: INFO: 1 became follower at term 0
raft2020/06/03 19:35:25 raft.go:383: INFO: newRaft 1 [peers: [1,2,3], term: 0, commit: 0, applied: 0, lastindex: 0, lastterm: 0]
raft2020/06/03 19:35:25 raft.go:900: INFO: 1 [term: 2] ignored a MsgApp message with lower term from 0 [term: 1]

一、section 5.2

  • 当服务器启动时,默认为follower状态
// TestStartAsFollower tests that when servers start up, they begin as followers.
// Reference: section 5.2
func TestStartAsFollower(t *testing.T) {
   r := newTestRaft(1, []uint64{1, 2, 3}, 10, 1, NewMemoryStorage())
   //r.state = StateCandidate
   if r.state != StateFollower {
      //如果不是follower将会报错
   	t.Errorf("state = %s, want %s", r.state, StateFollower)
   }
}
=== RUN   TestStartAsFollower
raft2020/06/04 11:48:18 raft.go:1521: INFO: 1 switched to configuration voters=(1 2 3)
raft2020/06/04 11:48:18 raft.go:699: INFO: 1 became follower at term 0
raft2020/06/04 11:48:18 raft.go:383: INFO: newRaft 1 [peers: [1,2,3], term: 0, commit: 0, applied: 0, lastindex: 0, lastterm: 0]
--- PASS: TestStartAsFollower (0.00s)
  • leader会定时发送心跳检测给其他follower以维护自己的地位
// TestLeaderBcastBeat tests that if the leader receives a heartbeat tick,
// it will send a MsgHeartbeat with m.Index = 0, m.LogTerm=0 and empty entries
// as heartbeat to all followers.
// Reference: section 5.2
func TestLeaderBcastBeat(t *testing.T) {
   // heartbeat interval
   hi := 1
   r := newTestRaft(1, []uint64{1, 2, 3}, 10, hi, NewMemoryStorage())  //10为选举超时时间,hi为心跳超时需要进行一次检测的时间
   r.becomeCandidate()
   r.becomeLeader()  //r.tick方法此时变为了tickHeartbeat;r.step方法变为stepLeader
   for i := 0; i < 10; i++ {
   	mustAppendEntry(r, pb.Entry{Index: uint64(i) + 1})
   }

   for i := 0; i < hi; i++ {
   	r.tick()  //发送一次心跳检测,运行tickHeartbeat方法;这里hi设为1所以只需滴答一次heartbeatElapsed会等于heartbeatTimeout
   }
   msgs := r.readMessages()  //该方法将r的message读出来,并且将r的message重置
   sort.Sort(messageSlice(msgs))
   wmsgs := []pb.Message{
   	{From: 1, To: 2, Term: 1, Type: pb.MsgHeartbeat},
   	{From: 1, To: 3, Term: 1, Type: pb.MsgHeartbeat},
   }
   if !reflect.DeepEqual(msgs, wmsgs) {
   	t.Errorf("msgs = %v, want %v", msgs, wmsgs)
   }
}

// tickHeartbeat is run by leaders to send a MsgBeat after r.heartbeatTimeout.
func (r *raft) tickHeartbeat() {
   r.heartbeatElapsed++  //每运行一遍该方法heartbeatElapsed增加1,当该值大于或等于heartbeatTimeout就会重置heartbeatElapsed并传递心跳检测
   r.electionElapsed++

   if r.electionElapsed >= r.electionTimeout {
   	r.electionElapsed = 0
   	if r.checkQuorum {
   		r.Step(pb.Message{From: r.id, Type: pb.MsgCheckQuorum})
   	}
   	// If current leader cannot transfer leadership in electionTimeout, it becomes leader again.
   	if r.state == StateLeader && r.leadTransferee != None {
   		r.abortLeaderTransfer()
   	}
   }

   if r.state != StateLeader {
   	return
   }

   if r.heartbeatElapsed >= r.heartbeatTimeout {
   	r.heartbeatElapsed = 0
   	r.Step(pb.Message{From: r.id, Type: pb.MsgBeat})
   	//传递心跳检测;m.Type为MsgBeat所以在Step方法里面跳到r.step即上面说的stepLeader;
   	//最后经过的方法:stepLeader->bcastHeartbeat->bcastHeartbeatWithCtx->sendHeartbeat->send
   	//最终在send里面将MsgBeat的信息添加到r.msg切片中去
   }
}
  • 重新选举的两种情况:1⃣️follower在选举超时时间内没收到心跳检测RPC,将会增加自己的当前任期并且转变为candidate状态,此时该ratf会先给自己投一票并发送vote RPC给其他servers。2⃣️集群中有多个candidate收到了相同选票,将会进入超时并且开始一次新的选举,任期号此生也增加一位
func TestFollowerStartElection(t *testing.T) {
   testNonleaderStartElection(t, StateFollower) 
}
func TestCandidateStartNewElection(t *testing.T) {
   testNonleaderStartElection(t, StateCandidate)
}

// testNonleaderStartElection tests that if a follower receives no communication
// over election timeout, it begins an election to choose a new leader. It
// increments its current term and transitions to candidate state. It then
// votes for itself and issues RequestVote RPCs in parallel to each of the
// other servers in the cluster.
// Reference: section 5.2
// Also if a candidate fails to obtain a majority, it will time out and
// start a new election by incrementing its term and initiating another
// round of RequestVote RPCs.
// Reference: section 5.2
func testNonleaderStartElection(t *testing.T, state StateType) {
   // election timeout
   et := 10 //选举超时时间10
   r := newTestRaft(1, []uint64{1, 2, 3}, et, 1, NewMemoryStorage())
   switch state {
   case StateFollower:
   	r.becomeFollower(1, 2)  //tick方法是tickElection
   case StateCandidate:
   	r.becomeCandidate()   //tick方法是tickElection
   }
//r会生成一个randomizedElectionTimeout,当下面的loop达到这个次数时,将进行MsgVote RPC
   for i := 1; i < 2*et; i++ {
   	r.tick() 
   }

   if r.Term != 2 {
   	t.Errorf("term = %d, want 2", r.Term)
   }
   if r.state != StateCandidate {
   	t.Errorf("state = %s, want %s", r.state, StateCandidate)
   }
   if !r.prs.Votes[r.id] {
   	t.Errorf("vote for self = false, want true")
   }
   msgs := r.readMessages()
   sort.Sort(messageSlice(msgs))
   wmsgs := []pb.Message{
   	{From: 1, To: 2, Term: 2, Type: pb.MsgVote},
   	{From: 1, To: 3, Term: 2, Type: pb.MsgVote},
   }
   if !reflect.DeepEqual(msgs, wmsgs) {
   	t.Errorf("msgs = %v, want %v", msgs, wmsgs)
   }
}
  • 当开始一次选举时,follower将会增加它的当前任期并且转变为cadidate状态。该状态将会持续到:1⃣️此节点赢了选举2⃣️别的节点赢的了选举并通知到位3⃣️一段时间后没有节点胜利
// TestLeaderElectionInOneRoundRPC tests all cases that may happen in
// leader election during one round of RequestVote RPC:
// a) it wins the election
// b) it loses the election
// c) it is unclear about the result
// Reference: section 5.2
func TestLeaderElectionInOneRoundRPC(t *testing.T) {
   tests := []struct {
   	size  int
   	votes map[uint64]bool
   	state StateType
   }{
   	// win the election when receiving votes from a majority of the servers
   	{1, map[uint64]bool{}, StateLeader},
   	{3, map[uint64]bool{2: true, 3: true}, StateLeader},
   	{3, map[uint64]bool{2: true}, StateLeader},
   	{5, map[uint64]bool{2: true, 3: true, 4: true, 5: true}, StateLeader},
   	{5, map[uint64]bool{2: true, 3: true, 4: true}, StateLeader},
   	{5, map[uint64]bool{2: true, 3: true}, StateLeader},

   	// return to follower state if it receives vote denial from a majority
   	{3, map[uint64]bool{2: false, 3: false}, StateFollower},
   	{5, map[uint64]bool{2: false, 3: false, 4: false, 5: false}, StateFollower},
   	{5, map[uint64]bool{2: true, 3: false, 4: false, 5: false}, StateFollower},

   	// stay in candidate if it does not obtain the majority
   	{3, map[uint64]bool{}, StateCandidate},
   	{5, map[uint64]bool{2: true}, StateCandidate},
   	{5, map[uint64]bool{2: false, 3: false}, StateCandidate},
   	{5, map[uint64]bool{}, StateCandidate},
   }
   for i, tt := range tests {
   	fmt.Println(i)
   	fmt.Println(tt)
   	r := newTestRaft(1, idsBySize(tt.size), 10, 1, NewMemoryStorage())

   	r.Step(pb.Message{From: 1, To: 1, Type: pb.MsgHup}) //对外发出投票请求
   	for id, vote := range tt.votes {
   		fmt.Println(id)
   		r.Step(pb.Message{From: id, To: 1, Term: r.Term, Type: pb.MsgVoteResp, Reject: !vote}) //模拟接收投票回来的结果,
   		//当发现投票超过了半数就becomeleader,当继续传回MsgVoteResp将不回有任何操作了
   	}

   	if r.state != tt.state {
   		t.Errorf("#%d: state = %s, want %s", i, r.state, tt.state)
   	}
   	if g := r.Term; g != 1 {
   		t.Errorf("#%d: term = %d, want %d", i, g, 1)
   	}
   }
}
  • 如果一个candidate在等待选票的时候,收到了一个追加日志的rpc,并且该rpc发出者的任期大于或等于该candidate,那么就认为这个发出追加日志rpc的leader是合法的。最后,该candidate会转变为follower
// TestCandidateFallback tests that while waiting for votes,
// if a candidate receives an AppendEntries RPC from another server claiming
// to be leader whose term is at least as large as the candidate's current term,
// it recognizes the leader as legitimate and returns to follower state.
// Reference: section 5.2
func TestCandidateFallback(t *testing.T) {
   tests := []pb.Message{
   	{From: 2, To: 1, Term: 1, Type: pb.MsgApp},
   	{From: 2, To: 1, Term: 2, Type: pb.MsgApp},
   }
   for i, tt := range tests {
   	r := newTestRaft(1, []uint64{1, 2, 3}, 10, 1, NewMemoryStorage())

   	r.Step(pb.Message{From: 1, To: 1, Type: pb.MsgHup})  //新建立的raft选举超时进行新的选举并向2和3发出投票RP,增加了一次任期
   	if r.state != StateCandidate {
   		t.Fatalf("unexpected state = %s, want %s", r.state, StateCandidate)
   	}

   	r.Step(tt)  //当回传的包里的任期参数大于当前raft时,1会变成follower;其他情况只要选票还没超过半数节点1还是处于candidate状态

   	if g := r.state; g != StateFollower {
   		t.Errorf("#%d: state = %s, want %s", i, g, StateFollower)
   	}
   	if g := r.Term; g != tt.Term {
   		t.Errorf("#%d: term = %d, want %d", i, g, tt.Term)
   	}
   }
}

待续

评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值