Zookeeper 源码分析-leader选举

选举的算法可以参考：http://blog.csdn.net/xhh198781/article/details/10949697

 

假设配置中有两个server

server.1=localhost:2888:3888

server.2=localhost:2889:3888

 

由前文可以，zookeeper在选举leader之前会先调用下面的代码，首先设置currentVote为myid，即一开始会选举自己为leader。如果electionType = 0，myid=1, Responder线程将监听在2888这个UDP端口上，处理其他节点的请求。

  synchronized public void startLeaderElection() {
        currentVote = new Vote(myid, getLastLoggedZxid());
        for (QuorumServer p : getView().values()) {
            if (p.id == myid) {
                myQuorumAddr = p.addr;
                break;
            }
        }
        if (myQuorumAddr == null) {
            throw new RuntimeException("My id " + myid + " not in the peer list");
        }
        if (electionType == 0) {
            try {
                udpSocket = new DatagramSocket(myQuorumAddr.getPort());
                responder = new ResponderThread();
                responder.start();
            } catch (SocketException e) {
                throw new RuntimeException(e);
            }
        }
        this.electionAlg = createElectionAlgorithm(electionType);
    }

 

 

LeaderElection

如果electionAlg=0，将使用LeaderElection算法。LeaderElection会调用lookForLeader方法，先对每个peer询问他们选举的leader，然后调用countVotes查看那个节点胜出，并将它设置成currentVote。如果超过半数的人选举这个节点，则选举成功。

  public Vote lookForLeader() throws InterruptedException {
            self.setCurrentVote(new Vote(self.getId(),
                    self.getLastLoggedZxid()));
         
            int xid = epochGen.nextInt();
            while (self.isRunning()) {
                requestBuffer.putInt(xid);
                requestPacket.setLength(4);
                HashSet heardFrom = new HashSet();
                for (QuorumServer server : self.getVotingView().values()) {
                    LOG.info("Server address: " + server.addr);
                    requestPacket.setSocketAddress(server.addr);
                    s.send(requestPacket);
                        responsePacket.setLength(responseBytes.length);
                        s.receive(responsePacket);
                        long peerId = responseBuffer.getLong();
                        heardFrom.add(peerId);
                        Vote vote = new Vote(responseBuffer.getLong(),
                                responseBuffer.getLong());
                        InetSocketAddress addr =
                                (InetSocketAddress) responsePacket
                                .getSocketAddress();
                        votes.put(addr, vote);
                }

                ElectionResult result = countVotes(votes, heardFrom);
                // ZOOKEEPER-569:
                // If no votes are received for live peers, reset to voting 
                // for ourselves as otherwise we may hang on to a vote 
                // for a dead peer                 
                if (votes.size() == 0) {                    
                    self.setCurrentVote(new Vote(self.getId(),
                            self.getLastLoggedZxid()));
                } else {
                    if (result.winner.id >= 0) {
                        self.setCurrentVote(result.vote);
                        // To do: this doesn't use a quorum verifier
                        if (result.winningCount > (self.getVotingView().size() / 2)) {
                            self.setCurrentVote(result.winner);
                             self.setPeerState((current.id == self.getId())
                                        ? ServerState.LEADING: ServerState.FOLLOWING);
                                if (self.getPeerState() == ServerState.FOLLOWING) {
                                    Thread.sleep(100);
                                }    								
                        }
					}
                }
			    
    }

 

 protected ElectionResult countVotes(HashMap votes, HashSet heardFrom) {
        ElectionResult result = new ElectionResult();
        // Initialize with null vote
        result.vote = new Vote(Long.MIN_VALUE, Long.MIN_VALUE);
        result.winner = new Vote(Long.MIN_VALUE, Long.MIN_VALUE);
        Collection votesCast = votes.values();
        // First make the views consistent. Sometimes peers will have
        // different zxids for a server depending on timing.
        for (Iterator i = votesCast.iterator(); i.hasNext();) {
            Vote v = i.next();
            if (!heardFrom.contains(v.id)) {
                // Discard votes for machines that we didn't hear from
                i.remove();
                continue;
            }
            for (Vote w : votesCast) {
                if (v.id == w.id) {
                    if (v.zxid < w.zxid) {
                        v.zxid = w.zxid;
                    }
                }
            }
        }

        HashMap countTable = new HashMap();
        // Now do the tally，选出zxid最大，且id最大的vote，作为下一轮选举的对象
        for (Vote v : votesCast) {
            Integer count = countTable.get(v);
            if (count == null) {
                count = Integer.valueOf(0);
            }
            countTable.put(v, count + 1);
            if (v.id == result.vote.id) {
                result.count++;
            } else if (v.zxid > result.vote.zxid
                    || (v.zxid == result.vote.zxid && v.id > result.vote.id)) {
                result.vote = v;
                result.count = 1;
            }
        }
        result.winningCount = 0;
        LOG.info("Election tally: ");
        //挑选出被选举最多的节点作为winner
        for (Entry entry : countTable.entrySet()) {
            if (entry.getValue() > result.winningCount) {
                result.winningCount = entry.getValue();
                result.winner = entry.getKey();
            }
            LOG.info(entry.getKey().id + "\t-> " + entry.getValue());
        }
        return result;
    }

 

    FastLeaderElection（另参考http://blog.csdn.net/xhh198781/article/details/6619203）

是标准的fast paxos的实现，它首先向所有Server提议自己要成为leader，当其它Server收到提议以后，解决 epoch 和 zxid 的冲突，并接受对方的提议，然后向对方发送接受提议完成的消息。

 

FastLeaderElection算法通过异步的通信方式来收集其它节点的选票，同时在分析选票时又根据投票者的当前状态来作不同的处理，以加快Leader的选举进程。

每个Server都一个接收线程池和一个发送线程池, 在没有发起选举时，这两个线程池处于阻塞状态，直到有消息到来时才解除阻塞并处理消息，同时每个Serve r都有一个选举线程(可以发起选举的线程担任)。

1). 主动发起选举端(选举线程)的处理

首先自己的 logicalclock加1，然后生成notification消息，并将消息放入发送队列中， 系统中配置有几个Server就生成几条消息，保证每个Server都能收到此消息，如果当前Server 的状态是LOOKING就一直循环检查接收队列是否有消息，如果有消息，根据消息中对方的状态进行相应的处理。

2).主动发送消息端(发送线程池)的处理

将要发送的消息由Notification消息转换成ToSend消息，然后发送对方，并等待对方的回复。

3). 被动接收消息端(接收线程池)的处理

将收到的消息转换成Notification消息放入接收队列中，如果对方Server的epoch小于logicalclock则向其发送一个消息(让其更新epoch)；如果对方Server处于Looking状态，自己则处于Following或Leading状态，则也发送一个消息(当前Leader已产生，让其尽快收敛)。

 /**
     * Starts a new round of leader election. Whenever our QuorumPeer
     * changes its state to LOOKING, this method is invoked, and it
     * sends notifications to all other peers.
     */
    public Vote lookForLeader() throws InterruptedException {
       HashMap recvset = new HashMap();
	    //发送notification给每个节点，告述他们自己当前选举的人，默认一开始选举自己   
		sendNotifications();

		/*
		 * Loop in which we exchange notifications until we find a leader
		 */

		while ((self.getPeerState() == ServerState.LOOKING) &&
				(!stop)){
			//当前节点启动了几个接收线程，用于接收其他节点选举的结果，并将选举的结果存到recvqueue中
			Notification n = recvqueue.poll(notTimeout,
					TimeUnit.MILLISECONDS);

			switch (n.state) {
				case LOOKING:
					// If notification > current, replace and send messages out
					if (n.epoch > logicalclock) {
						logicalclock = n.epoch;
						recvset.clear();
						//Check if a pair (server id, zxid) succeeds our current vote.
						if(totalOrderPredicate(n.leader, n.zxid,
								getInitId(), getInitLastLoggedZxid()))
							updateProposal(n.leader, n.zxid);
						else
							updateProposal(getInitId(),
									getInitLastLoggedZxid());
						sendNotifications();
					} else if (n.epoch < logicalclock) {
						if(LOG.isDebugEnabled()){
							LOG.debug("Notification epoch is smaller than logicalclock. n.epoch = " + n.epoch
									+ ", Logical clock" + logicalclock);
						}
						break;
					//如果n.zxid大于proposedZxid，或者n.zxid等于proposedZxid且leader id大于proposedLeader 的id, 那么选举n.leader
					} else if (totalOrderPredicate(n.leader, n.zxid,
							proposedLeader, proposedZxid)) {
						LOG.info("Updating proposal");
						updateProposal(n.leader, n.zxid);
						sendNotifications();
					}
 
					/*
					 * Only proceed if the vote comes from a replica in the
					 * voting view.
					 */
					if(self.getVotingView().containsKey(n.sid)){
						recvset.put(n.sid, new Vote(n.leader, n.zxid, n.epoch));

						//If have received from all nodes, then terminate
						if ((self.getVotingView().size() == recvset.size()) &&
								(self.getQuorumVerifier().getWeight(proposedLeader) != 0)){
							self.setPeerState((proposedLeader == self.getId()) ?
									ServerState.LEADING: learningState());
							leaveInstance();
							return new Vote(proposedLeader, proposedZxid);
						//如果满足结束的条件，将进入结束等待阶段
						} else if (termPredicate(recvset,
								new Vote(proposedLeader, proposedZxid,
										logicalclock))) {

							// Verify if there is any change in the proposed leader
							while((n = recvqueue.poll(finalizeWait,
									TimeUnit.MILLISECONDS)) != null){
								if(totalOrderPredicate(n.leader, n.zxid,
										proposedLeader, proposedZxid)){
									recvqueue.put(n);
									break;
								}
							}

							/*
							 * This predicate is true once we don't read any new
							 * relevant message from the reception queue
							 */
							if (n == null) {
								self.setPeerState((proposedLeader == self.getId()) ?
										ServerState.LEADING: learningState());
								if(LOG.isDebugEnabled()){
									LOG.debug("About to leave FLE instance: Leader= "
										+ proposedLeader + ", Zxid = " +
										proposedZxid + ", My id = " + self.getId()
										+ ", My state = " + self.getPeerState());
								}

								leaveInstance();
								return new Vote(proposedLeader,
										proposedZxid);
							}
						}
					}				
		}
    }