Github user mgummelt commented on a diff in the pull request:
https://github.com/apache/spark/pull/10993#discussion_r51927502
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/cluster/mesos/CoarseMesosSchedulerBackend.scala
---
@@ -373,40 +451,25 @@ private[spark] class CoarseMesosSchedulerBackend(
override def frameworkMessage(d: SchedulerDriver, e: ExecutorID, s:
SlaveID, b: Array[Byte]) {}
/**
- * Called when a slave is lost or a Mesos task finished. Update local
view on
- * what tasks are running and remove the terminated slave from the list
of pending
- * slave IDs that we might have asked to be killed. It also notifies the
driver
- * that an executor was removed.
+ * Called when a slave is lost or a Mesos task finished. Updates local
view on
+ * what tasks are running. It also notifies the driver that an executor
was removed.
*/
- private def executorTerminated(d: SchedulerDriver, slaveId: String,
reason: String): Unit = {
+ private def executorTerminated(d: SchedulerDriver,
+ slaveId: String,
+ taskId: String,
+ reason: String): Unit = {
stateLock.synchronized {
- if (slaveIdsWithExecutors.contains(slaveId)) {
- val slaveIdToTaskId = taskIdToSlaveId.inverse()
- if (slaveIdToTaskId.containsKey(slaveId)) {
- val taskId: Int = slaveIdToTaskId.get(slaveId)
- taskIdToSlaveId.remove(taskId)
- removeExecutor(sparkExecutorId(slaveId, taskId.toString),
SlaveLost(reason))
- }
- // TODO: This assumes one Spark executor per Mesos slave,
- // which may no longer be true after SPARK-5095
- pendingRemovedSlaveIds -= slaveId
- slaveIdsWithExecutors -= slaveId
- }
+ removeExecutor(taskId, SlaveLost(reason))
+ slaves(slaveId).taskIDs.remove(taskId)
--- End diff --
At the very least, this isn't a regression, because the previous SlaveID
HashMaps were not cleaned up either.
But even if we wanted to change that in this PR, how would we maintain the
current behaviour where the driver blacklists slaves and only registered with a
shuffle service once? We need `taskFailures` and `shuffleRegistered` state for
the lifetime of the driver.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
