kevin85421 commented on code in PR #37411:
URL: https://github.com/apache/spark/pull/37411#discussion_r943675732
##########
core/src/main/scala/org/apache/spark/HeartbeatReceiver.scala:
##########
@@ -199,41 +242,137 @@ private[spark] class HeartbeatReceiver(sc: SparkContext,
clock: Clock)
removeExecutor(executorRemoved.executorId)
}
+ private def killExecutor(executorId: String, timeout: Long): Unit = {
+ logWarning(s"Removing executor $executorId with no recent heartbeats: " +
+ s"${timeout} ms exceeds timeout $executorTimeoutMs ms")
+ killExecutorThread.submit(new Runnable {
+ override def run(): Unit = Utils.tryLogNonFatalError {
+ // Note: we want to get an executor back after expiring this one,
+ // so do not simply call `sc.killExecutor` here (SPARK-8119)
+ sc.killAndReplaceExecutor(executorId)
+ // SPARK-27348: in case of the executors which are not gracefully shut
down,
+ // we should remove lost executors from CoarseGrainedSchedulerBackend
manually
+ // here to guarantee two things:
+ // 1) explicitly remove executor information from
CoarseGrainedSchedulerBackend for
+ // a lost executor instead of waiting for disconnect message
+ // 2) call scheduler.executorLost() underlying to fail any tasks
assigned to
+ // those executors to avoid app hang
+ sc.schedulerBackend match {
+ case backend: CoarseGrainedSchedulerBackend =>
+ val isStandalone = backend.isInstanceOf[StandaloneSchedulerBackend]
+ backend.driverEndpoint.send(RemoveExecutor(executorId,
+ ExecutorProcessLost(
+ s"Executor heartbeat timed out after ${timeout} ms",
+ causedByApp = !checkWorkerLastHeartbeat || !isStandalone)))
Review Comment:
> For case (1), the task will keep getting rerun on a different executors -
where it will keep hanging repeatedly - if the hang is due to the user code.
You are right. To prevent infinite retries, we may temporarily set
`causedByApp` to true until we have a mechanism to detect hanging task in
executor. Does it make sense?
> You are right, (2) cannot be detected - but speculative execution might
help with that if it is related to the specific node (if it is related to user
code/data, it will hang indefinitely until user intervention).
As I mentioned above, we need a hanging task detector on executor to detect
this.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
