spark源码系列01-Woker启动Executor过程并向Driver注册_李孟_新浪博客

spark版本是1.3+

spark源码系列01-Woker启动Executor过程并向Driver注册

Woker启动Executor过程并向Driver注册时序图:
1.launchExecutor Master发送消息让Worker启动Executor

2.Worker new()
Master 发送给Worker的消息,让Worker启动Execitor,LaunchExecutor是一个Case Class,里面封装以后要启动的Executor的信息
new ExecutorRunner 创建ExcutorRunner,将参数都放到其中,然后在通过他启动Executor
注册ExecutorID -> Executor放到一个map中,对应关系
executors(appId + "/" + execId) = manager

3.Worker start()
调用ExecutorRunner的start方法来启动Executor Java子进程
manager.start() 

4.ExecutorRunner new()
先创建一个线程对象,然后通过一个线程来启动一个java子进程
workerThread = new Thread("ExecutorRunner for " + fullId) {
      override def run() { fetchAndRunExecutor() }
    }

5.ExecutorRunner 
workerThread.start()

6.ExecutorRunner 
程序经常也会遇到进程挂掉的情况,一些状态没有正确的保存下来,这时候就需要在JVM关掉的时候执行一些清理现场的代码 Runtime.getRuntime.addShutdownHook(shutdownHook)
shutdownHook = new Thread() {
      override def run() {
        killProcess(Some("Worker shutting down"))
      }
    }


7.ExecutorRunner fetchAndRunExecutor()
线程对象调用该方法启动java子进程

8.ExecutorRunner fetchAndRunExecutor()
保存参数
val builder = CommandUtils.buildProcessBuilder(appDesc.command, memory,
        sparkHome.getAbsolutePath, substituteVariables)


9.ExecutorRunner 
启动一个子进程 -> CoarseGrainedExecutorBackend的main方法
process = builder.start()


10.CoarseGrainedExecutorBackend main()
Executor进程执行的入口


11.CoarseGrainedExecutorBackend run()
调用RUN方法
run(driverUrl, executorId, hostname, cores, appId, workerUrl, userClassPath)

12.CoarseGrainedExecutorBackend preStart() CoarseGrainedExecutorBackend的生命周期方法
跟Driver建立连接
driver = context.actorSelection(driverUrl)

13.CoarseGrainedSchedulerBackend RegisterExecutor
Executor向DriverActor发送消息
 case RegisterExecutor(executorId, hostPort, cores, logUrls) ...

14.DriverActor CoarseGrainedExecutorBackend
DriverActor发送给Executor的消息,告诉他已经注册成功
case RegisteredExecutor => ...


15.DriverActor CoarseGrainedSchedulerBackend
查看是否有任务需要提交(DriverActor -> Executor)
makeOffers()

16.CoarseGrainedExecutorBackend new()
创建一个Executor实例,用来执行业务逻辑
executor = new Executor(executorId, hostname, env, userClassPath, isLocal = false)


17.Executor 
初始化线程池
val threadPool = Utils.newDaemonCachedThreadPool("Executor task launch worker")

18.Executor
Executor向DriverActor发送心跳,是否存活
startDriverHeartbeater()
posted @ 2018-04-26 12:28  Dlimeng  阅读(14)  评论(0)    收藏  举报  来源