核心枚举
public enum ServerState { LOOKING, FOLLOWING, LEADING, OBSERVING; }
zookeeper服务器状态:刚启动LOOKING,follower是FOLLOWING,leader是LEADING,observer是OBSERVING;
public enum LearnerType { PARTICIPANT, OBSERVER; }
简单来说,zookeeper启动的核心类是QuorumPeerMain,启动之后会加载配置,同时启动QuorumPeer,QuorumPeer会从本地数据目录加载数据,然后开始选举,选举的核心类是FastLeaderElection,选举完成之后会设置状态,根据状态的不同,启动不同的类,比如Leader、Follower、Observer,leader会设置一个新的epoch,同时会等待follower向自己注册,注册完成后follower会向leader同步最新的数据,而leader则定期ping所有的follower,一旦存活的follower不足半数,leader会shutdown,然后触发集群的重新选举。
具体代码以及调用如下:
org.apache.zookeeper.server.quorum.QuorumPeerMain
initializeAndRun
runFromConfig
org.apache.zookeeper.server.quorum.QuorumPeer extends ZooKeeperThread
initialize
start
loadDataBase
org.apache.zookeeper.server.ServerCnxnFactory
start
startLeaderElection
org.apache.zookeeper.server.quorum.QuorumPeer.ResponderThread
start
createElectionAlgorithm
run
org.apache.zookeeper.server.quorum.FastLeaderElection implements Election
lookForLeader
sendNotifications
totalOrderPredicate
termPredicate
org.apache.zookeeper.server.quorum.flexible.QuorumMaj
containsQuorum
QuorumPeer.setPeerState
1 org.apache.zookeeper.server.quorum.Leader extends Learner
lead
org.apache.zookeeper.server.quorum.Leader.LearnerCnxAcceptor extends ZooKeeperThread
run
org.apache.zookeeper.server.quorum.LearnerHandler extends ZooKeeperThread
start
receive Leader.FOLLOWERINFO
epoch++
setZxid
setCurrentEpoch
waitForNewLeaderAck
startZkServer
loop
org.apache.zookeeper.server.quorum.LearnerHandler
ping
org.apache.zookeeper.server.quorum.LearnerHandler.SyncLimitCheck
check
shutdown
Leader.removeLearnerHandler
containsQuorum
shutdown
2 org.apache.zookeeper.server.quorum.Follower extends Learner
followLeader
findLeader
connectToLeader
registerWithLeader
send Leader.FOLLOWERINFO
syncWithLeader
3 org.apache.zookeeper.server.quorum.Observer extends Learner
observeLeader
选举代码解析请见 大数据基础之Zookeeper(3)选举算法