背景
本文基于SPARK 3.3.0
在Spark 3.3.0中出现了一个新特性那就是自动重启Executor,这个主要解决是什么问题呢? 主要解决在Streaming中由于一个Executor的处理延迟导致整个Streaming任务延迟,但是这也是适用于批任务,使得批任务Executor的驱逐更加灵活。具体的可参考SPARK-37810
分析
在spark 3.3.0之前,如果发现任务是比较慢活着任务失败了,
可以开启spark.speculation(默认是关闭的),进行推测执行,
也可以开启spark.excludeOnFailure.enabled (默认是关闭的)以保证task不会重新调度到失败的Executor上
如果发现executor失败了,可以开启spark.excludeOnFailure.killExcludedExecutors(默认是关闭的),确保在fetch失败的时候,把execlude给删除掉。
但是这些都是事后的弥补方式,所以这里提出的Executor Rolling是事前预测执行的方式,该方式会周期性的轮询。
直接看代码ExecutorRollPlugin:
class ExecutorRollPlugin extends SparkPlugin { override def driverPlugin(): DriverPlugin = new ExecutorRollDriverPlugin() // No-op override def executorPlugin(): ExecutorPlugin = null }
它继承SparkPlugin,关于SparkPlugin,可以参考spark 3.x Plugin Framework,总的来说,spark提供了一种插件机制,我们可以灵活的用它来做自己想要的事情,比如说 自定义指标等等。
我们看到这里executorPlugin方法是为null的,因为Executor的启动停止调度是在Driver进行的,所以executor根本不需要。
而对于ExecutorRollDriverPlugin:
class ExecutorRollDriverPlugin extends DriverPlugin with Logging { override def init(sc: SparkContext, ctx: PluginContext): JMap[String, String] = { val interval = sc.conf.get(EXECUTOR_ROLL_INTERVAL) if (interval <= 0) { logWarning(s"Disabled due to invalid interval value, '$interval'") } else if (!sc.conf.get(DECOMMISSION_ENABLED)) { logWarning(s"Disabled because ${DECOMMISSION_ENABLED.key} is false.") } else { minTasks = sc.conf.get(MINIMUM_TASKS_PER_EXECUTOR_BEFORE_ROLLING) // Scheduler is not created yet sparkContext = sc val policy = ExecutorRollPolicy.withName(sc.conf.get(EXECUTOR_ROLL_POLICY)) periodicService.scheduleAtFixedRate(() => { try { sparkContext.schedulerBackend match { case scheduler: KubernetesClusterSchedulerBackend => val executorSummaryList = sparkContext .statusStore .executorList(true) choose(executorSummaryList, policy) match { case Some(id) => // Use decommission to be safe. logInfo(s"Ask to decommission executor $id") val now = System.currentTimeMillis() scheduler.decommissionExecutor( id, ExecutorDecommissionInfo(s"Rolling via $policy at $now"), adjustTargetNumExecutors = false) case _ => logInfo("There is nothing to roll.") } case _ => logWarning("This plugin expects " + s"${classOf[KubernetesClusterSchedulerBackend].getSimpleName}.") } } catch { case e: Throwable => logError("Error in rolling thread", e) } }, interval, interval, TimeUnit.SECONDS) } Map.empty[String, String].asJava } .... private def choose(list: Seq[v1.ExecutorSummary], policy: ExecutorRollPolicy.Value) : Option[String] = { val listWithoutDriver = list .filterNot(_.id.equals(SparkContext.DRIVER_IDENTIFIER)) .filter(_.totalTasks >= minTasks) val sortedList = policy match { case ExecutorRollPolicy.ID => // We can convert to integer because EXECUTOR_ID_COUNTER uses AtomicInteger. listWithoutDriver.sortBy(_.id.toInt) case ExecutorRollPolicy.ADD_TIME => listWithoutDriver.sortBy(_.addTime) case ExecutorRollPolicy.TOTAL_GC_TIME => listWithoutDriver.sortBy(_.totalGCTime).reverse case ExecutorRollPolicy.TOTAL_DURATION => listWithoutDriver.sortBy(_.totalDuration).reverse case ExecutorRollPolicy.AVERAGE_DURATION => listWithoutDriver.sortBy(e => e.totalDuration.toFloat / Math.max(1, e.totalTasks)).reverse case ExecutorRollPolicy.FAILED_TASKS => listWithoutDriver.sortBy(_.failedTasks).reverse case ExecutorRollPolicy.OUTLIER => // If there is no outlier we fallback to TOTAL_DURATION policy. outliersFromMultipleDimensions(listWithoutDriver) ++ listWithoutDriver.sortBy(_.totalDuration).reverse case ExecutorRollPolicy.OUTLIER_NO_FALLBACK => outliersFromMultipleDimensions(listWithoutDriver) } sortedList.headOption.map(_.id) }
这里是periodicService单个线程定时触发,如果发现backend是k8s的话(所以目前只适用于spark on k8s),就会从已有的AppStatusStore(通过AppStatusListener机制获取到对应的Event,从而存储信息,目前来看,executor的metrics信息是通过heartbeat来传递到driver端的)存储中取出Executor的信息,进而根据配置的策略(Executor创建的ID,失败的task,GC时间等)进行驱逐。
当然在驱逐Executor的时候,也会考虑目前在Executor上运行的task的个数,具体配置为spark.kubernetes.executor.minTasksPerExecutorBeforeRolling(默认是0),只有小于等于该阈值,才会kill 对应的Executor,而且默认是只驱逐一个Executor。
该方式的优点:
- 事前的处理方式,而不是事后处理
- 独立于ExecutorPodsAllocator,使组件之间功能明确,便于代码维护
- 适用于动态和静态Executor资源分配的场景
- 驱逐策略根据运行时的统计信息来的,更加合理
具体使用方式,配置如下:
spark.plugins=org.apache.spark.scheduler.cluster.k8s.ExecutorRollPlugin spark.decommission.enabled=true spark.kubernetes.executor.rollInterval=3600s