LiveListenerBus

官方说明如下:

Asynchronously passes SparkListenerEvents to registered SparkListeners.

即它的功能是异步地将SparkListenerEvent传递给已经注册的SparkListener,这种异步的机制是通过生产消费者模型来实现的。

首先,它定义了 4 个 消息堵塞队列,队列的名字分别为shared、appStatus、executorManagement、eventLog。队列的类型是 org.apache.spark.scheduler.AsyncEventQueue#AsyncEventQueue,保存在 queues 变量中。每一个队列上都可以注册监听器,如果队列没有监听器,则会被移除。

它有启动和stop和start两个标志位来指示 监听总线的的启动停止状态。 如果总线没有启动,有事件过来,先放到 一个待添加的可变数组中,否则直接将事件 post 到每一个队列中。

其直接依赖类是 AsyncEventQueue, 相当于 LiveListenerBus 的多事件队列是对 AsyncEventQueue 进一步的封装。

AsyncEventQueue

其继承关系如下:

  spark 源码分析之三 -- LiveListenerBus介绍-LMLPHP

它有启动和stop和start两个标志位来指示 监听总线的的启动停止状态。

其内部维护了listenersPlusTimers 主要就是用来保存注册到这个总线上的监听器对象的。

post 操作将事件放入内部的 LinkedBlockingQueue中,默认大小是 10000。

有一个事件分发器,它不停地从 LinkedBlockingQueue 执行 take 操作,获取事件,并将事件进一步分发给所有的监听器,由org.apache.spark.scheduler.SparkListenerBus#doPostEvent 方法实现事件转发,具体代码如下:

 1 protected override def doPostEvent(
2 listener: SparkListenerInterface,
3 event: SparkListenerEvent): Unit = {
4 event match {
5 case stageSubmitted: SparkListenerStageSubmitted =>
6 listener.onStageSubmitted(stageSubmitted)
7 case stageCompleted: SparkListenerStageCompleted =>
8 listener.onStageCompleted(stageCompleted)
9 case jobStart: SparkListenerJobStart =>
10 listener.onJobStart(jobStart)
11 case jobEnd: SparkListenerJobEnd =>
12 listener.onJobEnd(jobEnd)
13 case taskStart: SparkListenerTaskStart =>
14 listener.onTaskStart(taskStart)
15 case taskGettingResult: SparkListenerTaskGettingResult =>
16 listener.onTaskGettingResult(taskGettingResult)
17 case taskEnd: SparkListenerTaskEnd =>
18 listener.onTaskEnd(taskEnd)
19 case environmentUpdate: SparkListenerEnvironmentUpdate =>
20 listener.onEnvironmentUpdate(environmentUpdate)
21 case blockManagerAdded: SparkListenerBlockManagerAdded =>
22 listener.onBlockManagerAdded(blockManagerAdded)
23 case blockManagerRemoved: SparkListenerBlockManagerRemoved =>
24 listener.onBlockManagerRemoved(blockManagerRemoved)
25 case unpersistRDD: SparkListenerUnpersistRDD =>
26 listener.onUnpersistRDD(unpersistRDD)
27 case applicationStart: SparkListenerApplicationStart =>
28 listener.onApplicationStart(applicationStart)
29 case applicationEnd: SparkListenerApplicationEnd =>
30 listener.onApplicationEnd(applicationEnd)
31 case metricsUpdate: SparkListenerExecutorMetricsUpdate =>
32 listener.onExecutorMetricsUpdate(metricsUpdate)
33 case executorAdded: SparkListenerExecutorAdded =>
34 listener.onExecutorAdded(executorAdded)
35 case executorRemoved: SparkListenerExecutorRemoved =>
36 listener.onExecutorRemoved(executorRemoved)
37 case executorBlacklistedForStage: SparkListenerExecutorBlacklistedForStage =>
38 listener.onExecutorBlacklistedForStage(executorBlacklistedForStage)
39 case nodeBlacklistedForStage: SparkListenerNodeBlacklistedForStage =>
40 listener.onNodeBlacklistedForStage(nodeBlacklistedForStage)
41 case executorBlacklisted: SparkListenerExecutorBlacklisted =>
42 listener.onExecutorBlacklisted(executorBlacklisted)
43 case executorUnblacklisted: SparkListenerExecutorUnblacklisted =>
44 listener.onExecutorUnblacklisted(executorUnblacklisted)
45 case nodeBlacklisted: SparkListenerNodeBlacklisted =>
46 listener.onNodeBlacklisted(nodeBlacklisted)
47 case nodeUnblacklisted: SparkListenerNodeUnblacklisted =>
48 listener.onNodeUnblacklisted(nodeUnblacklisted)
49 case blockUpdated: SparkListenerBlockUpdated =>
50 listener.onBlockUpdated(blockUpdated)
51 case speculativeTaskSubmitted: SparkListenerSpeculativeTaskSubmitted =>
52 listener.onSpeculativeTaskSubmitted(speculativeTaskSubmitted)
53 case _ => listener.onOtherEvent(event)
54 }
55 }

然后去调用 listener 的相对应的方法。

就这样,事件总线上的消息事件被监听器消费了。

05-29 01:04