【Hadoop】完全分布式下Resourcemanager二次分发及Job提交到集群后状态的变化

用于查看Hadoop中完全分布式下Resourcemanager二次分发以及作业提交到集群后状态的变化。

IDE：eclipse

调试IDE所在操作系统：win10

Hadoop集群操作系统：Ubuntu16.04

Hadoop集群环境：完全分布式模式，版本为2.7.3

1. 两个概念

IPC : 进程间通信
RPC : 远程过程调用

2. EventHandler

EventHandler<T extends Event>       // 用于处理和它关联的事件Event
Event<TYPE extends Enum<TYPE>>      // 每个事件Event都有与其关联的类型TYPE
TYPE      // 事件类型TYPE，即事件拥有的各种状态，即事件生命周期中的各个过程，用于状态机对其状态的转换

3. RM二次分发图解

4. 二次分发调试解析

ResourceManager

--> ResourceManager.main()                                      // ResourceManager入口main函数
    --> new YarnConfiguration()                                 // 设置YarnConfiguration，即core-default.xml, core-site.xml, yarn-default.xml, yarn-site.xml
    --> resourceManager.init(conf)                              // 初始化resourceManager
        --> AbstractService.init()                              // 调用父类AbstractService的初始化init方法
            --> ResourceManager.serviceInit(config)             // 调用ResourceManager自身的serviceInit方法，即回调过程
                --> conf.addResource(coreSiteXMLInputStream)    // 加载core-site.xml
                --> conf.addResource(yarnSiteXMLInputStream)    // 加载yarn-site.xml
                --> rmContext.setHAEnabled(...)                 // 设置HA
                --> setupDispatcher()                           // 设置分发器，register the handlers for all AlwaysOn services
                    --> createDispatcher()                      // 创建分发器
                        --> new AsyncDispatcher()               // 用构造新建分发器
                --> addService(adminService)                    // 为服务注册处理器，即把映射关系放入AsyncDispatcher.eventDispatchers集合中
                --> add...                                      // 为其他常在服务注册处理器

AsyncDispatcher

--> serviceStart()                                              // 用于在事件状态由INITED向STARTED转换时调用，即事件初始化完成，启动事件时调用
    --> createThread()                                          // 创建提取分发线程
        --> eventQueue.take()                                   // 从事件队列中提取事件
        --> dispatch(event)                                     // 按照事件类型一次分发事件
            --> event.getType().getDeclaringClass()             // 获取事件类型
            --> eventDispatchers.get(type)                      // 根据类型获取相应的处理器
            --> handler.handle(event)                           // 调用处理器的handle方法进行处理
                --> rmContext.getRMApps().get(appID)            // 通过appID获取RMApp
                --> rmApp.handle(event)                         // 调用RMApp的handle方法进行处理，即二次分发

5. job提交到集群之后的状态变换

在集群上提交一个job，断点停在ResourceManager$ApplicationEventDispatcher.handle方法
--> handle(RMAppEvent event)                                    // EventType: START || UI State: NEW
    --> event.getApplicationId()                                // ApplicationId: application_1489305385892_0001
    --> rmContext.getRMApps().get(appID)                        // 获取RMApp: org.apache.hadoop.yarn.server.resourcemanager.rmapp.RMAppImpl@252ec2a1
                                                                // Dispatcher: STARTED
    --> RMAppImpl.handle(event)                                 // 交由RMAppImpl处理，即二次分发
        --> RMAppState oldState = getState()                    // 获取之前一次状态: NEW
        --> stateMachine.doTransition(event.getType(), event)   // 状态机进行状态变换
--> // 事件转换状态之后，仍会被放到eventQueue队列中等待处理
--> AsyncDispatcher$GenericEventHandler.handle(Event event)

--> handle(RMAppEvent event)                                    // EventType: APP_NEW_SAVED         || UI State: NEW_SAVING
    --> ...[重复以上过程]

--> handle(RMAppEvent event)                                    // EventType: APP_ACCEPTED          || UI State: SUBMITTED
    --> ...

--> ApplicationAttemptEventDispatcher.handle(RMAppEvent event)  // EventType: NEW                   || UI State: ACCEPTED
    --> ...

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: START                 || UI State: ACCEPTED
    --> ...

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: ATTEMPT_ADDED         || UI State: ACCEPTED
    --> ...

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: CONTAINER_ALLOCATED   || UI State: ACCEPTED
    --> ...

--> handle(RMAppEvent event)                                    // EventType: APP_RUNNING_ON_NODE   || UI State: ACCEPTED
    --> // 更新进度

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: ATTEMPT_NEW_SAVED     || UI State: ACCEPTED
    --> ...

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: LAUNCHED              || UI State: ACCEPTED
    --> ...

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: REGISTERED            || UI State: ACCEPTED
    --> ...

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: STATUS_UPDATE         || UI State: ACCEPTED [尝试多次]
    --> ...

--> handle(RMAppEvent event)                                    // EventType: APP_RUNNING_ON_NODE   || UI State: RUNNING
    --> // 更新进度

--> ApplicationAttemptEventDispatcherhandle(RMAppEvent event)   // EventType: STATUS_UPDATE         || UI State: RUNNING [尝试多次]
    --> ...

--> handle(RMAppEvent event)                                    // EventType: APP_RUNNING_ON_NODE   || UI State: RUNNING
    --> // 更新进度

--> handle(RMAppEvent event)                                    // EventType: ATTEMPT_UNREGISTERED  || UI State: RUNNING
    --> // 更新进度

--> handle(RMAppEvent event)                                    // EventType: APP_UPDATE_SAVED      || UI State: RUNNING    || FinalStatus: SUCCEEDED
    --> // 更新进度

--> FINISHED

　　其中可能会出现一些问题，而导致进入ResourceManager$ApplicationAttemptEventDispatcher.handle方法中进行尝试，尝试的过程中，UI界面就是ACCEPTED状态。
　　从ApplicationEventDispatcher分发器可以看出Application在集群上的状态变换： START -> APP_NEW_SAVED -> APP_ACCEPTED -> APP_RUNNING_ON_NODE[多次处于该状态] -> ATTEMPT_UNREGISTERED -> APP_UPDATE_SAVED
　　webUI中显示出的Application状态变换：NEW -> NEW_SAVING -> SUBMITTED -> ACCEPTED -> RUNNING -> FINISHED