一、简介

      This document gives an overview of NodeManager (NM) restart, a feature that enables the NodeManager to be restarted without losing the active containers running on the node. At a high level, the NM stores any necessary state to a local state-store as it processes container-management requests. When the NM restarts, it recovers by first loading state for various subsystems and then letting those subsystems perform recovery using the loaded state.

      这个是官网的介绍。NodeManager Restart Recover是Hadoop-2.6.0全面引入的一个新特性。它旨在实现NodeManager在不丢失Active Containers、Localized Resource、Applications等的情况下的重启，NodeManager会使用Leveldb记录Containers、Localized Resource、Applications等的关键请求状态，在NodeManager重启后实现状态恢复等。

二、开启

      如果要开启NodeManager重启特性，需要配置以下两个参数：

  <property>
    <description>Enable the node manager to recover after starting</description>
    <name>yarn.nodemanager.recovery.enabled</name>
    <value>true</value>
  </property>
  
  <property>
    <description>The local filesystem directory in which the node manager will
    store state when recovery is enabled.</description>
    <name>yarn.nodemanager.recovery.dir</name>
    <value>${hadoop.tmp.dir}/yarn-nm-recovery</value>
  </property>

三、实现范围

       NodeManager Restart Recover的实现范围如下：

      1、Localized Resource State Storage and Recovery
      2、Application State Storage and Recovery
      3、Container State Storage and Recovery
      4、NM Token State Storage and Recovery
      5、Container Token State Storage and Recovery
      6、Log Aggregation Recovery
      7、Auxiliary Service State Storage and Recovery 

四、实现分析

      1、NMStateStoreService引入

      首先，在NodeManager中，有一个NMStateStoreService组件，定义如下：

  private NMStateStoreService nmStore = null;

      这个NMStateStoreService组件就是NodeManager Restart Recover的具体实现，它实现了Containers、Localized Resource、Applications等关键事件状态的存储、恢复、删除的等；

      NMStateStoreService是一个抽象类，继承自AbstractService这个抽象服务，并在服务的serviceInit()、serviceStart()、serviceStop()等方法中分别调用了存储相关的initStorage(conf)、startStorage()、closeStorage()三个方法，分别完成存储的初始化、开始、结束等工作。

      以Applications为例，NMStateStoreService中定义了如下几个抽象方法：

  /**
   * Load the state of applications
   * @return recovered state for applications
   * @throws IOException
   */
  public abstract RecoveredApplicationsState loadApplicationsState()
      throws IOException;

  /**
   * Record the start of an application
   * @param appId the application ID
   * @param p state to store for the application
   * @throws IOException
   */
  public abstract void storeApplication(ApplicationId appId,
      ContainerManagerApplicationProto p) throws IOException;

  /**
   * Record that an application has finished
   * @param appId the application ID
   * @throws IOException
   */
  public abstract void storeFinishedApplication(ApplicationId appId)
      throws IOException;

  /**
   * Remove records corresponding to an application
   * @param appId the application ID
   * @throws IOException
   */
  public abstract void removeApplication(ApplicationId appId)
      throws IOException;

      这几个方法分别实现了应用的存储、已完成应用的存储、应用的删除、应用的加载（即恢复）等。大体流程是这样的，应用在NodeManager上初始化后即调用storeApplication()方法存储应用的状态（方便以后恢复），当ResourceManager指示Application完成后，storeFinishedApplication()方法被调用，以表明该Application已完成，而当NodeManager不再需要追踪该Application时，removeApplication()被调用以删除该Application的状态信息。最后，如果NodeManager需要重启，loadApplicationsState()方法被调用以实现需要恢复状态的Applications的状态恢复。

      2、NMStateStoreService实现

      NMStateStoreService组件的实现目前有两种：NMNullStateStoreService、NMLeveldbStateStoreService，而NMNullStateStoreService仅仅是空的NMStateStoreService组件，里面的上述关键方法，如initStorage(conf)、startStorage()、closeStorage()、storeApplication()等方法全部是空方法，而NMLeveldbStateStoreService则是上述参数配置后实现NodeManager Restart Recover的具体实现组件。

      3、NMStateStoreService初始化、启动与停止

      在NodeManager的serviceInit()方法中，有如下调用：

initAndStartRecoveryStore(conf);

      而initAndStartRecoveryStore()方法中，则会判断上述两个参数，完成NMStateStoreService的初始化，如下：

  private void initAndStartRecoveryStore(Configuration conf)
      throws IOException {
    boolean recoveryEnabled = conf.getBoolean(
        YarnConfiguration.NM_RECOVERY_ENABLED,
        YarnConfiguration.DEFAULT_NM_RECOVERY_ENABLED);
    if (recoveryEnabled) {
      FileSystem recoveryFs = FileSystem.getLocal(conf);
      String recoveryDirName = conf.get(YarnConfiguration.NM_RECOVERY_DIR);
      if (recoveryDirName == null) {
        throw new IllegalArgumentException("Recovery is enabled but " +
            YarnConfiguration.NM_RECOVERY_DIR + " is not set.");
      }
      Path recoveryRoot = new Path(recoveryDirName);
      recoveryFs.mkdirs(recoveryRoot, new FsPermission((short)0700));
      nmStore = new NMLeveldbStateStoreService();
    } else {
      nmStore = new NMNullStateStoreService();
    }
    nmStore.init(conf);
    nmStore.start();
  }

      当参数yarn.nodemanager.recovery.enabled配置成true时，NMStateStoreService初始化为NMLeveldbStateStoreService，并初始化存储路径yarn.nodemanager.recovery.dir，否则为NMNullStateStoreService。

      继而调用NMStateStoreService的init()和start()方法，完成初始化和启动。

      最后，当NodeManager服务停止时，在其serviceStop()方法中会调用stopRecoveryStore()方法，其会调用NMStateStoreService的stop()方法，停止存储服务。

      4、NMLeveldbStateStoreService实现分析