ZooKeeper源码阅读（五）：Leader选举

ZooKeeper中的Leader选举也不是Paxos, 实现相关的类包括FastLeaderElection, LeaderElection, 继承层次如下:

//Election 

//  +AuthFastLeaderElection

//  +FastLeaderElection

//    +MockFLE (测试)

//  +LeaderElection

//    +MockLeaderElection

 

默认使用的是FastLeaderElection, 不过ZooKeeper其实并不关心Leader选举是如何实现的,只要满足:

• The leader has seen the highest zxid of all the followers.

• A quorum of servers have committed to following the leader.

如果有follower的zxid比Leader看到的zxid还大,说明是Leader选举结束之后才连上Leader的.这时Leader会发送TRUNC消息,让Follower丢弃.

1) FastLeaderElection的实现

 

//构造函数

       //两个队列，在Messenger中将会用到

     sendqueue = newLinkedBlockingQueue<ToSend>();

     recvqueue = newLinkedBlockingQueue<Notification>();

 

//messenger对象

//Messenger启动了两个线程：WorkerSender，WorkerReceiver

     this.messenger =newMessenger(manager);

                  

//WorkerSender消费sendqueue

         ToSendm = sendqueue.poll(3000,TimeUnit.MILLISECONDS);

       process(m);

 

//队列中消息类型为ToSend

//{leader, zxid,elecEpoch, state, sid, peerEpoch, configData}

         //表示的消息为Notification或者对Notification的回复

 

//WorkerReceiver接收消息

         //如果不是参与投票的，直接回复自己的投票，放到sendqueue中

         //

         //否则，解析成Notification。自己还是Looking则放到recvqueue中，

         //自己不是Looking，对方是Looking，则回复自己的leader编号

 

//lookForLeader开始新的一轮选举

         //更新logicalclock

         synchronized(this){

            logicalclock++;

           updateProposal(getInitId(), getInitLastLoggedZxid(), getPeerEpoch());

        }

 

       sendNotifications();

 

//发送Notification，直到不是LOOKING

        while ((self.getPeerState()== ServerState.LOOKING) &&

             (!stop)){

            /*

             *Remove next notification from queue, times out after 2 times

             * thetermination time

             */

           Notification n=recvqueue.poll(notTimeout, TimeUnit.MILLISECONDS);

 

                   //如果是投票者(VoterFollower)，根据它的状态

                   //LOOKING:

                   //更新logicclock,更新Proposal，发送Notification

                   //放入recvset，然后判断是否达到多数派

                   //如果达到，则等待finalizeWait微秒，没有更优的投票则选举结束

 

                   //对方是Following,Leading

                   //如果logicalclock相符，检查是否达到多数派

                   //放到outofelection的都是LEADING或者FOLLOWING状态的，已经选举完毕

                  //都要checkLeader(outofelection,n.leader, n.electionEpoch)检查

 

                   //checkLeader

        /*

         * Ifeveryone else thinks I'm the leader, I must be the leader.

         * Theother two checks are just for the case in which I'm not the

         * leader.If I'm not the leader and I haven't received a message

         * fromleader stating that it is leading, then predicate is false.

         */

 

        if(leader !=self.getId()){

            if(votes.get(leader)==null) predicate = false;

            else if(votes.get(leader).getState()!= ServerState.LEADING) predicate =false;

        } else if(logicalclock !=electionEpoch) {

           predicate = false;

        }

 

logicalclock的增加表示新一轮的选举过程.

Leader选举初始投票选自己,收到别人的投票之后判断是否更优,如果是则更新自己的投票,最终zxid最大的follower将收到多数派的投票,等待finalizeWait微秒后仍无更优投票,则转为LEADING状态.

 

2) LeaderElection的实现则更简单

         //向所有VotingView中的Server建立socket，发送/接收vote

         //然后countVotes判断截止

3) AuthFastLeaderElection则是在FastLeaderElection基础上增加了简单的认证

 

Leader选举时，将选择zxid最高的。这也是一个优化，因为不用和其它组员那里找丢失的事务。

参考:

http://zookeeper.apache.org/doc/r3.2.2/zookeeperInternals.html#sc_leaderElection