1. SparkContext提供了一个取消job的api
class SparkContext(config: SparkConf) extends Logging with ExecutorAllocationClient { /** Cancel a given job if it's scheduled or running */ private[spark] def cancelJob(jobId: Int) { dagScheduler.cancelJob(jobId) } }
2. 那么如何获取jobId呢?
Spark提供了一个叫SparkListener的对象,它提供了对spark事件的监听功能
trait SparkListener { /** * Called when a job starts */ def onJobStart(jobStart: SparkListenerJobStart) { } /** * Called when a job ends */ def onJobEnd(jobEnd: SparkListenerJobEnd) { } }
因此需要自定义一个类,继承自SparkListener,即:
public class DHSparkListener implements SparkListener { private static Logger logger = Logger.getLogger(DHSparkListener.class); //存储了提交job的线程局部变量和job的映射关系 private static ConcurrentHashMap<String, Integer> jobInfoMap; public DHSparkListener() { jobInfoMap = new ConcurrentHashMap<String, Integer>(); } @Override public void onJobEnd(SparkListenerJobEnd jobEnd) { logger.info("DHSparkListener Job End:" + jobEnd.jobResult().getClass() + ",Id:" + jobEnd.jobId()); for (String key : jobInfoMap.keySet()) { if (jobInfoMap.get(key) == jobEnd.jobId()) { jobInfoMap.remove(key); logger.info(key+" request has been returned. because "+jobEnd.jobResult().getClass()); } } } @Override public void onJobStart(SparkListenerJobStart jobStart) { logger.info("DHSparkListener Job Start: JobId->" + jobStart.jobId()); //根据线程变量属性找到该job是哪个线程提交的 logger.info("DHSparkListener Job Start: Thread->" + jobStart.properties().getProperty("thread", "default")); jobInfoMap.put(jobStart.properties().getProperty("thread", "default"), jobStart.jobId()); } …… }
那么用户如何知道该job是哪个线程提交的呢?需要在提交job的时候设置线程局部变量属性,即
SparkConf conf = new SparkConf().setAppName("SparkListenerTest application in Java"); String sparkMaster = Configure.instance.get("SparkMaster"); String sparkExecutorMemory = "16g"; String sparkCoresMax = "4"; String sparkJarAddress = "/tmp/cuckoo-core-1.0-SNAPSHOT-allinone.jar"; conf.setMaster(sparkMaster); conf.set("spark.executor.memory", sparkExecutorMemory); conf.set("spark.cores.max", sparkCoresMax); JavaSparkContext jsc = new JavaSparkContext(conf); jsc.addJar(sparkJarAddress); DHSparkListener dHSparkListener = new DHSparkListener(); jsc.sc().addSparkListener(dHSparkListener); List<Integer> listData = new ArrayList<Integer>(); listData = Arrays.asList(1, 2, 3, 4, 5, 6, 7, 8, 9); JavaRDD<Integer> rdd1 = jsc.parallelize(listData, 1); JavaRDD<Integer> rdd2 = rdd1.map(new Function<Integer, Integer>() { public Integer call(Integer v1) throws Exception { //do something then return } }); <pre name="code" class="plain"> //在触发action提交job之前设置提交线程的局部属性,供SparkListener获取 jsc.setLocalProperty("thread", "client"); rdd2.count();
这样在jobInfoMap中记录了job和job提交者的映射关系,当发现某个job迟迟没有结束的时候,可以调用SparkContext的cancelJob取消,但是仅仅到这里就够了吗?接着往下看,excutor取消job最终调用的是:
def kill(interruptThread: Boolean) { _killed = true if (context != null) { context.markInterrupted() } if (interruptThread && taskThread != null) { taskThread.interrupt() } }
最终调用到Thread.interrupt函数,给启动task的线程设置interrupt标记位,因此在长时间允许的task中,需要针对Thread的interrupt标记位进行判断,当被置位的时候,需要退出,并且做一些清理,即存在类似的代码段:
if(Thread.interrupted()){ //……线程被中断,清理资源 } 或者调用sleep,wait函数时会抛出InterruptedException异常,需要进行捕获,然后做对应的处理
3. 最后一步,配置job kill的动作
除了以上操作之外,还需要再配置针对每个job调用kill的动作,即spark.job.interruptOnCancel属性为true
//在触发action提交job之前设置提交线程的局部属性,供SparkListener获取 jsc.setLocalProperty("thread", "client"); //配置该job接受到kill之后的动作,即task线程收到interrupt信号 jsc.setLocalProperty("spark.job.interruptOnCancel", "true"); rdd2.count();