ResourceManager 内维护了 NodeManager 的生命周期；对于每个 NodeManager 在 ResourceManager 中都有一个 RMNode 与其对应；除了 RMNode ，ResourceManager 中还定义了 NodeManager 的状态（states）以及触发状态转移的事件（event）。具体如下：

RMNodeImpl 中定义了 NodeManager 的状态转移情况，如下：

private static final StateMachineFactory<RMNodeImpl,
                                         NodeState,
                                         RMNodeEventType,
                                         RMNodeEvent> stateMachineFactory 
               = new StateMachineFactory<RMNodeImpl,
                                         NodeState,
                                         RMNodeEventType,
                                         RMNodeEvent>(NodeState.NEW)

   //Transitions from NEW state
   .addTransition(NodeState.NEW, NodeState.RUNNING, 
       RMNodeEventType.STARTED, new AddNodeTransition())

   //Transitions from RUNNING state
   .addTransition(NodeState.RUNNING, 
       EnumSet.of(NodeState.RUNNING, NodeState.UNHEALTHY),
       RMNodeEventType.STATUS_UPDATE, new StatusUpdateWhenHealthyTransition())
   .addTransition(NodeState.RUNNING, NodeState.DECOMMISSIONED,
       RMNodeEventType.DECOMMISSION,
       new DeactivateNodeTransition(NodeState.DECOMMISSIONED))
   .addTransition(NodeState.RUNNING, NodeState.LOST,
       RMNodeEventType.EXPIRE,
       new DeactivateNodeTransition(NodeState.LOST))
   .addTransition(NodeState.RUNNING, NodeState.REBOOTED,
       RMNodeEventType.REBOOTING,
       new DeactivateNodeTransition(NodeState.REBOOTED))
   .addTransition(NodeState.RUNNING, NodeState.RUNNING,
       RMNodeEventType.CLEANUP_APP, new CleanUpAppTransition())
   .addTransition(NodeState.RUNNING, NodeState.RUNNING,
       RMNodeEventType.CLEANUP_CONTAINER, new CleanUpContainerTransition())
   .addTransition(NodeState.RUNNING, NodeState.RUNNING,
       RMNodeEventType.RECONNECTED, new ReconnectNodeTransition())

   //Transitions from UNHEALTHY state
   .addTransition(NodeState.UNHEALTHY, 
       EnumSet.of(NodeState.UNHEALTHY, NodeState.RUNNING),
       RMNodeEventType.STATUS_UPDATE, new StatusUpdateWhenUnHealthyTransition())
   .addTransition(NodeState.UNHEALTHY, NodeState.DECOMMISSIONED,
       RMNodeEventType.DECOMMISSION,
       new DeactivateNodeTransition(NodeState.DECOMMISSIONED))
   .addTransition(NodeState.UNHEALTHY, NodeState.LOST,
       RMNodeEventType.EXPIRE,
       new DeactivateNodeTransition(NodeState.LOST))
   .addTransition(NodeState.UNHEALTHY, NodeState.REBOOTED,
       RMNodeEventType.REBOOTING,
       new DeactivateNodeTransition(NodeState.REBOOTED))
   .addTransition(NodeState.UNHEALTHY, NodeState.UNHEALTHY,
       RMNodeEventType.RECONNECTED, new ReconnectNodeTransition())
   .addTransition(NodeState.UNHEALTHY, NodeState.UNHEALTHY,
       RMNodeEventType.CLEANUP_APP, new CleanUpAppTransition())
   .addTransition(NodeState.UNHEALTHY, NodeState.UNHEALTHY,
       RMNodeEventType.CLEANUP_CONTAINER, new CleanUpContainerTransition())
       
   // create the topology tables
   .installTopology(); 

上面定义了 NodeManager 每种状态之间的转移及其触发的事件。比如从 NodeState.NEW 状态转移到 NodeState.RUNNIN 的触发事件为 RMNodeEventType.STARTED，其状态转移处理类为 AddNodeTransition 。根据上面的定义我们可以画出 NodeManager 整个生命周期转换图：

如果想及时了解Spark、Hadoop或者Hbase相关的文章，欢迎关注微信公共帐号：iteblog_hadoop

从上图可以看出，NodeManager 生命周期的开始状态和最终状态：

所有的 NodeManager 向 ResourceManager 注册的初始化状态都为 NodeState.NEW，一旦向 ResourceManager 注册成功其状态会更新为 NodeState.RUNNIN。还会处理以下两件事：

NodeManager 在启动之后会默认每隔 1s （由 yarn.resourcemanager.nodemanagers.heartbeat-interval-ms 参数配置）向 ResourceManager 发送心跳信息。在 ResourceManager 端会启动 PingChecker 线程默认每隔 200s （yarn.nm.liveness-monitor.expiry-interval-ms 参数值得三分之一）检测所有注册到 ResourceManager 的节点，一旦发现有节点超过 600s （由 yarn.nm.liveness-monitor.expiry-interval-ms 参数配置）没有发送心跳信息，则认为这个节点出问题了，这时候将会把这个节点从 running 列表中移除，并发送 RMNodeEventType.EXPIRE 事件。RMNodeImpl 接收到这个事件会通过 DeactivateNodeTransition 类来处理，这个类会将此节点从 ResourceManager 可用节点列表中移除，并加入到 inactiveNodes 列表里面，此时 NodeManager 状态由 NodeState.RUNNING 或 NodeState.UNHEALTHY 状态转移到 NodeState.LOST 状态，而且该节点被标记为 NodesListManagerEventType.NODE_UNUSABLE。所有在这个节点上运行的 containers 将会标记为失败，并且会分配到新的 NodeManager 上运行。

NodeManager 其他状态之间的转移及其触发的事件这里就不再详细阐述。关于 NodeManager 自身的状况监测机制，请点击这里。