Github user pwendell commented on a diff in the pull request:
https://github.com/apache/spark/pull/1056#discussion_r15682056
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala ---
@@ -320,6 +323,26 @@ private[spark] class TaskSchedulerImpl(
}
}
+ /**
+ * Update metrics for in-progress tasks and let the master know that the
BlockManager is still
+ * alive. Return true if the driver knows about the given block manager.
Otherwise, return false,
+ * indicating that the block manager should re-register.
+ */
+ override def executorHeartbeatReceived(
+ execId: String,
+ taskMetrics: Array[(Long, TaskMetrics)], // taskId -> TaskMetrics
+ blockManagerId: BlockManagerId): Boolean = {
+ val metricsWithStageIds = taskMetrics.flatMap {
+ case (id, metrics) => {
+ taskIdToTaskSetId.get(id)
--- End diff --
I think there is an unlikely race here where (a) a task heartbeat gets
enqueued to be sent (b) the task actually finishes and that message is sent,
then taskIdToTaskSetId is cleared (c) the heartbeat arrives. This is possible
since the heartbeater and the task execution itself are in different threads.
Then you'd get an NPE here. Though extremely unlikely, it might be good to just
log a warning and pass if the task set is not found.
---
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.
---
