Spark技术内幕:Master基于ZooKeeper的HighAvailability(HA)源码实现

  • 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HighAvailability(HA)源码实现_第1张图片


    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HAFILESYSTEM: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。

    view source print ?
    01. override def preStart() {
    02.     logInfo("Starting Spark master at " + masterUrl)
    03.     ...
    04.     //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
    05.     //已经提交Job的运行
    06.     persistenceEngine = RECOVERY_MODE match {
    07.       case "ZOOKEEPER" =>
    08.         logInfo("Persisting recovery state to ZooKeeper")
    09.         new ZooKeeperPersistenceEngine(SerializationExtension(context.system), conf)
    10.       case "FILESYSTEM" =>
    11.         logInfo("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。

    view source print ?
    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。

    view source print ?
    01. private[spark] class BlackHolePersistenceEngine extends PersistenceEngine {
    02.   override def addApplication(app: ApplicationInfo) {}
    03.   override def removeApplication(app: ApplicationInfo) {}
    04.   override def addWorker(worker: WorkerInfo) {}
    05.   override def removeWorker(worker: WorkerInfo) {}
    06.   override def addDriver(driver: DriverInfo) {}
    07.   override def removeDriver(driver: DriverInfo) {}
    08.   
    09.   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。

    view source print ?
    01. class ZooKeeperPersistenceEngine(serialization: Serialization, conf: SparkConf)
    02.   extends PersistenceEngine
    03.   with Logging
    04. {
    05.   val WORKING_DIR = conf.get("spark.deploy.zookeeper.dir", "/spark") + "/master_status"
    06.   val zk: CuratorFramework = SparkCuratorUtil.newClient(conf)
    07.   
    08.   SparkCuratorUtil.mkdir(zk, WORKING_DIR)
    09.   // 将app的信息序列化到文件WORKING_DIR/app_{app.id}中
    10.   override def addApplication(app: ApplicationInfo) {
    11.     serializeIntoFile(WORKING_DIR + "/app_" + app.id, app)
    12.   }
    13.   
    14.   override def removeApplication(app: ApplicationInfo) {
    15.     zk.delete().forPath(WORKING_DIR + "/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:

    view source print ?
    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()。

view source print ?
01. /**
02.      * Create a new client
03.      *
04.      *
05.      * @param connectString list of servers to connect to
06.      * @param sessionTimeoutMs session timeout
07.      * @param connectionTimeoutMs connection timeout
08.      * @param retryPolicy retry policy to use
09.      * @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。

view source print ?
01. /**
02. * LeaderLatchListener can be used to be notified asynchronously about when the state of the LeaderLatch has changed.
03. *
04. * Note that just because you are in the middle of one of these method calls, it does not necessarily mean that
05. * hasLeadership() is the corresponding true/false value. It is possible for the state to change behind the scenes
06. * before these methods get called. The contract is that if that happens, you should see another call to the other
07. * method pretty quickly.
08. */
09. 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

view source print ?
01. public class LeaderLatch implements Closeable
02. {
03.     private final Logger log = LoggerFactory.getLogger(getClass());
04.     private final CuratorFramework client;
05.     private final String latchPath;
06.     private final String id;
07.     private final AtomicReference<State> state = new AtomicReference<State>(State.LATENT);
08.     private final AtomicBoolean hasLeadership = new AtomicBoolean(false);
09.     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。

view source print ?
01. private[spark] class ZooKeeperLeaderElectionAgent(val masterActor: ActorRef,
02.     masterUrl: String, conf: SparkConf)
03.   extends LeaderElectionAgent with LeaderLatchListener with Logging  {
04.   val WORKING_DIR = conf.get("spark.deploy.zookeeper.dir", "/spark") + "/leader_election"
05.   // zk是通过CuratorFrameworkFactory创建的ZooKeeper实例
06.   private var zk: CuratorFramework = _
07.   // leaderLatch:Curator负责选出Leader。
08.   private var leaderLatch: LeaderLatch = _
09.   private var status = LeadershipStatus.NOT_LEADER
10.   
11.   override def preStart() {
12.   
13.     logInfo("Starting ZooKeeper LeaderElection agent")
14.     zk = SparkCuratorUtil.newClient(conf)
15.     leaderLatch = new LeaderLatch(zk, WORKING_DIR)
16.     leaderLatch.addListener(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。

view source print ?
01. override def isLeader() {
02.   synchronized {
03.     // could have lost leadership by now.
04.     //现在leadership可能已经被剥夺了。。详情参见Curator的实现。
05.     if (!leaderLatch.hasLeadership) {
06.       return
07.     }
08.     logInfo("We have gained leadership")
09.     updateLeadershipStatus(true)
10.   }
11. }
12. override def notLeader() {
13.   synchronized {
14.     // 现在可能赋予leadership了。详情参见Curator的实现。
15.     if (leaderLatch.hasLeadership) {
16.       return
17.     }
18.     logInfo("We have lost leadership")
19.     updateLeadershipStatus(false)
20.   }
21. }

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

view source print ?
01. def updateLeadershipStatus(isLeader: Boolean) {
02.     if (isLeader && status == LeadershipStatus.NOT_LEADER) {
03.       status = LeadershipStatus.LEADER
04.       masterActor ! ElectedLeader
05.     } else if (!isLeader && status == LeadershipStatus.LEADER) {
06.       status = LeadershipStatus.NOT_LEADER
07.       masterActor ! RevokedLeadership
08.     }
09.   }

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?

延伸阅读:

  • 1、nginx源码学习Unix - Unix域协议
  • 2、tornado源码分析系列(一)
  • 3、tornado源码分析系列(二)
  • 4、Easyui1.32源码翻译--datagrid(数据表格)
  • 5、Hadoop源码分析之IPC连接与方法调用
  • 6、Hadoop源码分析之客户端读取HDFS数据
  • 7、AsyncTask源码分析
  • 8、openfire3.9.1源码部署及运行

你可能感兴趣的:(Spark技术内幕:Master基于ZooKeeper的HighAvailability(HA)源码实现)