Spark技术内幕：Master基于ZooKeeper的High Availability（HA）源码实现-CFANZ编程社区

如果Spark的部署方式选择Standalone，一个采用Master/Slaves的典型架构，那么Master是有SPOF（单点故障，Single Point of Failure）。Spark可以选用ZooKeeper来实现HA。

ZooKeeper提供了一个Leader Election机制，利用这个机制可以保证虽然集群存在多个Master但是只有一个是Active的，其他的都是Standby，当Active的Master出现故障时，另外的一个Standby Master会被选举出来。由于集群的信息，包括Worker， Driver和Application的信息都已经持久化到文件系统，因此在切换的过程中只会影响新Job的提交，对于正在进行的Job没有任何的影响。加入ZooKeeper的集群整体架构如下图所示。

Spark技术内幕：Master基于ZooKeeper的High Availability（HA）源码实现_spark

1. Master的重启策略

Master在启动时，会根据启动参数来决定不同的Master故障重启策略：

ZOOKEEPER实现HA
FILESYSTEM：实现Master无数据丢失重启，集群的运行时数据会保存到本地/网络文件系统上
丢弃所有原来的数据重启

Master::preStart()可以看出这三种不同逻辑的实现。

1. override def preStart() {  
2. "Starting Spark master at " + masterUrl)  
3.     ...  
4. //persistenceEngine是持久化Worker，Driver和Application信息的，这样在Master重新启动时不会影响  
5. //已经提交Job的运行  
6.     persistenceEngine = RECOVERY_MODE match {  
7. case "ZOOKEEPER" =>  
8. "Persisting recovery state to ZooKeeper")  
9. new ZooKeeperPersistenceEngine(SerializationExtension(context.system), conf)  
10. case "FILESYSTEM" =>  
11. "Persisting recovery state to directory: " + RECOVERY_DIR)  
12. new FileSystemPersistenceEngine(RECOVERY_DIR, SerializationExtension(context.system))  
13. case _ =>  
14. new BlackHolePersistenceEngine()  
15.     }  
16. //leaderElectionAgent负责Leader的选取。  
17.     leaderElectionAgent = RECOVERY_MODE match {  
18. case "ZOOKEEPER" =>  
19.           context.actorOf(Props(classOf[ZooKeeperLeaderElectionAgent], self, masterUrl, conf))  
20. case _ => // 仅仅有一个Master的集群，那么当前的Master就是Active的  
21.           context.actorOf(Props(classOf[MonarchyLeaderAgent], self))  
22.       }  
23.   }

RECOVERY_MODE是一个字符串，可以从spark-env.sh中去设置。

1. val RECOVERY_MODE = conf.get("spark.deploy.recoveryMode", "NONE")

如果不设置spark.deploy.recoveryMode的话，那么集群的所有运行数据在Master重启是都会丢失，这个结论是从BlackHolePersistenceEngine的实现得出的。

1. private[spark] class BlackHolePersistenceEngine extends PersistenceEngine {  
2.   override def addApplication(app: ApplicationInfo) {}  
3.   override def removeApplication(app: ApplicationInfo) {}  
4.   override def addWorker(worker: WorkerInfo) {}  
5.   override def removeWorker(worker: WorkerInfo) {}  
6.   override def addDriver(driver: DriverInfo) {}  
7.   override def removeDriver(driver: DriverInfo) {}  
8.   
9.   override def readPersistedData() = (Nil, Nil, Nil)  
10. }

它把所有的接口实现为空。PersistenceEngine是一个trait。作为对比，可以看一下ZooKeeper的实现。

1. class ZooKeeperPersistenceEngine(serialization: Serialization, conf: SparkConf)  
2. extends PersistenceEngine  
3.   with Logging  
4. {  
5. "spark.deploy.zookeeper.dir", "/spark") + "/master_status"  
6.   val zk: CuratorFramework = SparkCuratorUtil.newClient(conf)  
7.   
8.   SparkCuratorUtil.mkdir(zk, WORKING_DIR)  
9. // 将app的信息序列化到文件WORKING_DIR/app_{app.id}中  
10.   override def addApplication(app: ApplicationInfo) {  
11. "/app_" + app.id, app)  
12.   }  
13.   
14.   override def removeApplication(app: ApplicationInfo) {  
15. "/app_" + app.id)  
16.   }

Spark使用的并不是ZooKeeper的API，而是使用的org.apache.curator.framework.CuratorFramework 和 org.apache.curator.framework.recipes.leader.{LeaderLatchListener, LeaderLatch} 。Curator在ZooKeeper上做了一层很友好的封装。

2. 集群启动参数的配置

简单总结一下参数的设置，通过上述代码的分析，我们知道为了使用ZooKeeper至少应该设置一下参数（实际上，仅仅需要设置这些参数。通过设置spark-env.sh：

1. spark.deploy.recoveryMode=ZOOKEEPER  
2. spark.deploy.zookeeper.url=zk_server_1:2181,zk_server_2:2181  
3. spark.deploy.zookeeper.dir=/dir     
4. // OR 通过一下方式设置  
5. export SPARK_DAEMON_JAVA_OPTS="-Dspark.deploy.recoveryMode=ZOOKEEPER "  
6. export SPARK_DAEMON_JAVA_OPTS="${SPARK_DAEMON_JAVA_OPTS} -Dspark.deploy.zookeeper.url=zk_server1:2181,zk_server_2:2181"

各个参数的意义：

参数	默认值	含义
spark.deploy.recoveryMode	NONE	恢复模式（Master重新启动的模式），有三种：1, ZooKeeper, 2， FileSystem, 3 NONE
spark.deploy.zookeeper.url		ZooKeeper的Server地址
spark.deploy.zookeeper.dir	/spark	ZooKeeper 保存集群元数据信息的文件目录，包括Worker，Driver和Application。

3. CuratorFramework简介

CuratorFramework极大的简化了ZooKeeper的使用，它提供了high-level的API，并且基于ZooKeeper添加了很多特性，包括

自动连接管理：连接到ZooKeeper的Client有可能会连接中断，Curator处理了这种情况，对于Client来说自动重连是透明的。
简洁的API：简化了原生态的ZooKeeper的方法，事件等；提供了一个简单易用的接口。
Recipe的实现（更多介绍请点击Recipes）：

Leader的选择
共享锁
缓存和监控
分布式的队列
分布式的优先队列

CuratorFrameworks通过CuratorFrameworkFactory来创建线程安全的ZooKeeper的实例。

CuratorFrameworkFactory.newClient()提供了一个简单的方式来创建ZooKeeper的实例，可以传入不同的参数来对实例进行完全的控制。获取实例后，必须通过start()来启动这个实例，在结束时，需要调用close()。

1. /**
2.      * Create a new client
3.      *
4.      *
5.      * @param connectString list of servers to connect to
6.      * @param sessionTimeoutMs session timeout
7.      * @param connectionTimeoutMs connection timeout
8.      * @param retryPolicy retry policy to use
9.      * @return client
10.      */  
11. public static CuratorFramework newClient(String connectString, int sessionTimeoutMs, int connectionTimeoutMs, RetryPolicy retryPolicy)  
12.     {  
13. return builder().  
14.             connectString(connectString).  
15.             sessionTimeoutMs(sessionTimeoutMs).  
16.             connectionTimeoutMs(connectionTimeoutMs).  
17.             retryPolicy(retryPolicy).  
18.             build();  
19.     }

需要关注的还有两个Recipe：org.apache.curator.framework.recipes.leader.{LeaderLatchListener, LeaderLatch}。

首先看一下LeaderlatchListener，它在LeaderLatch状态变化的时候被通知：

在该节点被选为Leader的时候，接口isLeader()会被调用
在节点被剥夺Leader的时候，接口notLeader()会被调用

由于通知是异步的，因此有可能在接口被调用的时候，这个状态是准确的，需要确认一下LeaderLatch的hasLeadership()是否的确是true/false。这一点在接下来Spark的实现中可以得到体现。

1. /**
2. * LeaderLatchListener can be used to be notified asynchronously about when the state of the LeaderLatch has changed.
3. *
4. * Note that just because you are in the middle of one of these method calls, it does not necessarily mean that
5. * hasLeadership() is the corresponding true/false value. It is possible for the state to change behind the scenes
6. * before these methods get called. The contract is that if that happens, you should see another call to the other
7. * method pretty quickly.
8. */  
9. public interface LeaderLatchListener  
10. {  
11. /**
12. * This is called when the LeaderLatch's state goes from hasLeadership = false to hasLeadership = true.
13. *
14. * Note that it is possible that by the time this method call happens, hasLeadership has fallen back to false. If
15. * this occurs, you can expect {@link #notLeader()} to also be called.
16. */  
17. public void isLeader();  
18.   
19. /**
20. * This is called when the LeaderLatch's state goes from hasLeadership = true to hasLeadership = false.
21. *
22. * Note that it is possible that by the time this method call happens, hasLeadership has become true. If
23. * this occurs, you can expect {@link #isLeader()} to also be called.
24. */  
25. public void notLeader();  
26. }

LeaderLatch负责在众多连接到ZooKeeper Cluster的竞争者中选择一个Leader。Leader的选择机制可以看ZooKeeper的具体实现，LeaderLatch这是完成了很好的封装。我们只需要要知道在初始化它的实例后，需要通过

1. public class LeaderLatch implements Closeable  
2. {  
3. private final Logger log = LoggerFactory.getLogger(getClass());  
4. private final CuratorFramework client;  
5. private final String latchPath;  
6. private final String id;  
7. private final AtomicReference<State> state = new AtomicReference<State>(State.LATENT);  
8. private final AtomicBoolean hasLeadership = new AtomicBoolean(false);  
9. private final AtomicReference<String> ourPath = new AtomicReference<String>();  
10. private final ListenerContainer<LeaderLatchListener> listeners = new ListenerContainer<LeaderLatchListener>();  
11. private final CloseMode closeMode;  
12. private final AtomicReference<Future<?>> startTask = new AtomicReference<Future<?>>();  
13. .  
14. .  
15. .  
16. /**
17.      * Attaches a listener to this LeaderLatch
18.      * <p/>
19.      * Attaching the same listener multiple times is a noop from the second time on.
20.      * <p/>
21.      * All methods for the listener are run using the provided Executor.  It is common to pass in a single-threaded
22.      * executor so that you can be certain that listener methods are called in sequence, but if you are fine with
23.      * them being called out of order you are welcome to use multiple threads.
24.      *
25.      * @param listener the listener to attach
26.      */  
27. public void addListener(LeaderLatchListener listener)  
28.     {  
29.         listeners.addListener(listener);  
30.     }

通过addListener可以将我们实现的Listener添加到LeaderLatch。在Listener里，我们在两个接口里实现了被选为Leader或者被剥夺Leader角色时的逻辑即可。

4. ZooKeeperLeaderElectionAgent的实现

实际上因为有Curator的存在，Spark实现Master的HA就变得非常简单了，ZooKeeperLeaderElectionAgent实现了接口LeaderLatchListener，在isLeader()确认所属的Master被选为Leader后，向Master发送消息ElectedLeader，Master会将自己的状态改为ALIVE。当noLeader()被调用时，它会向Master发送消息RevokedLeadership时，Master会关闭。

1. private[spark] class ZooKeeperLeaderElectionAgent(val masterActor: ActorRef,  
2.     masterUrl: String, conf: SparkConf)  
3. extends LeaderElectionAgent with LeaderLatchListener with Logging  {  
4. "spark.deploy.zookeeper.dir", "/spark") + "/leader_election"  
5. // zk是通过CuratorFrameworkFactory创建的ZooKeeper实例  
6. private var zk: CuratorFramework = _  
7. // leaderLatch：Curator负责选出Leader。  
8. private var leaderLatch: LeaderLatch = _  
9. private var status = LeadershipStatus.NOT_LEADER  
10.   
11.   override def preStart() {  
12.   
13. "Starting ZooKeeper LeaderElection agent")  
14.     zk = SparkCuratorUtil.newClient(conf)  
15. new LeaderLatch(zk, WORKING_DIR)  
16. this)  
17.   
18.     leaderLatch.start()  
19.   }

在prestart中，启动了leaderLatch来处理选举ZK中的Leader。就如在上节分析的，主要的逻辑在isLeader和noLeader中。

1. override def isLeader() {  
2. synchronized {  
3. // could have lost leadership by now.  
4. //现在leadership可能已经被剥夺了。。详情参见Curator的实现。  
5. if (!leaderLatch.hasLeadership) {  
6. return  
7.     }  
8.   
9. "We have gained leadership")  
10. true)  
11.   }  
12. }  
13.   
14. override def notLeader() {  
15. synchronized {  
16. // 现在可能赋予leadership了。详情参见Curator的实现。  
17. if (leaderLatch.hasLeadership) {  
18. return  
19.     }  
20.   
21. "We have lost leadership")  
22. false)  
23.   }  
24. }

updateLeadershipStatus的逻辑很简单，就是向Master发送消息。

1. def updateLeadershipStatus(isLeader: Boolean) {  
2. if (isLeader && status == LeadershipStatus.NOT_LEADER) {  
3.       status = LeadershipStatus.LEADER  
4.       masterActor ! ElectedLeader  
5. else if (!isLeader && status == LeadershipStatus.LEADER) {  
6.       status = LeadershipStatus.NOT_LEADER  
7.       masterActor ! RevokedLeadership  
8.     }  
9.   }

5. 设计理念

为了解决Standalone模式下的Master的SPOF，Spark采用了ZooKeeper提供的选举功能。Spark并没有采用ZooKeeper原生的Java API，而是采用了Curator，一个对ZooKeeper进行了封装的框架。采用了Curator后，Spark不用管理与ZooKeeper的连接，这些对于Spark来说都是透明的。Spark仅仅使用了100行代码，就实现了Master的HA。当然了，Spark是站在的巨人的肩膀上。谁又会去重复发明轮子呢？