[ https://issues.apache.org/jira/browse/SPARK-4006?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Tal Sliwowicz updated SPARK-4006: --------------------------------- Description: This is a huge robustness issue for us, in mission critical , time sensitive (real time) spark jobs. We have long running spark drivers and even though we have state of the art hardware, from time to time executors disconnect. In many cases, the RemoveExecutor is not received, and when the new executor registers, the driver crashes. In mesos coarse grained, executor ids are fixed. The issue is with the System.exit(1) in BlockManagerMasterActor private def register(id: BlockManagerId, maxMemSize: Long, slaveActor: ActorRef) { if (!blockManagerInfo.contains(id)) { blockManagerIdByExecutor.get(id.executorId) match { case Some(manager) => // A block manager of the same executor already exists. // This should never happen. Let's just quit. logError("Got two different block manager registrations on " + id.executorId) System.exit(1) case None => blockManagerIdByExecutor(id.executorId) = id } logInfo("Registering block manager %s with %s RAM".format( id.hostPort, Utils.bytesToString(maxMemSize))) blockManagerInfo(id) = new BlockManagerInfo(id, System.currentTimeMillis(), maxMemSize, slaveActor) } listenerBus.post(SparkListenerBlockManagerAdded(id, maxMemSize)) } was: We have long running spark drivers and even though we have state of the art hardware, from time to time executors disconnect. In many cases, the RemoveExecutor is not received, and when the new executor registers, the driver crashes. In mesos coarse grained, executor ids are fixed. The issue is with the System.exit(1) in BlockManagerMasterActor private def register(id: BlockManagerId, maxMemSize: Long, slaveActor: ActorRef) { if (!blockManagerInfo.contains(id)) { blockManagerIdByExecutor.get(id.executorId) match { case Some(manager) => // A block manager of the same executor already exists. // This should never happen. Let's just quit. logError("Got two different block manager registrations on " + id.executorId) System.exit(1) case None => blockManagerIdByExecutor(id.executorId) = id } logInfo("Registering block manager %s with %s RAM".format( id.hostPort, Utils.bytesToString(maxMemSize))) blockManagerInfo(id) = new BlockManagerInfo(id, System.currentTimeMillis(), maxMemSize, slaveActor) } listenerBus.post(SparkListenerBlockManagerAdded(id, maxMemSize)) } > Spark Driver crashes whenever an Executor is registered twice > ------------------------------------------------------------- > > Key: SPARK-4006 > URL: https://issues.apache.org/jira/browse/SPARK-4006 > Project: Spark > Issue Type: Bug > Components: Block Manager, Spark Core > Affects Versions: 0.9.2, 1.0.2, 1.1.0 > Environment: Mesos, Coarse Grained > Reporter: Tal Sliwowicz > Priority: Critical > > This is a huge robustness issue for us, in mission critical , time sensitive > (real time) spark jobs. > We have long running spark drivers and even though we have state of the art > hardware, from time to time executors disconnect. In many cases, the > RemoveExecutor is not received, and when the new executor registers, the > driver crashes. In mesos coarse grained, executor ids are fixed. > The issue is with the System.exit(1) in BlockManagerMasterActor > private def register(id: BlockManagerId, maxMemSize: Long, slaveActor: > ActorRef) { > if (!blockManagerInfo.contains(id)) { > blockManagerIdByExecutor.get(id.executorId) match { > case Some(manager) => > // A block manager of the same executor already exists. > // This should never happen. Let's just quit. > logError("Got two different block manager registrations on " + > id.executorId) > System.exit(1) > case None => > blockManagerIdByExecutor(id.executorId) = id > } > logInfo("Registering block manager %s with %s RAM".format( > id.hostPort, Utils.bytesToString(maxMemSize))) > blockManagerInfo(id) = > new BlockManagerInfo(id, System.currentTimeMillis(), maxMemSize, > slaveActor) > } > listenerBus.post(SparkListenerBlockManagerAdded(id, maxMemSize)) > } -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org