venkata91 commented on a change in pull request #28287:
URL: https://github.com/apache/spark/pull/28287#discussion_r448697727



##########
File path: core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala
##########
@@ -289,13 +290,23 @@ private[spark] class ExecutorAllocationManager(
       s" tasksperexecutor: $tasksPerExecutor")
     val maxNeeded = math.ceil(numRunningOrPendingTasks * 
executorAllocationRatio /
       tasksPerExecutor).toInt
-    if (tasksPerExecutor > 1 && maxNeeded == 1 && pendingSpeculative > 0) {
+    val totalNeed = if (tasksPerExecutor > 1 && maxNeeded == 1 && 
pendingSpeculative > 0) {
       // If we have pending speculative tasks and only need a single executor, 
allocate one more
       // to satisfy the locality requirements of speculation
       maxNeeded + 1
     } else {
       maxNeeded
     }
+
+    // Request additional executors to schedule the unschedulable tasks as well
+    if (numUnschedulables > 0) {
+      val maxNeededForUnschedulables = math.ceil(numUnschedulables * 
executorAllocationRatio /
+        tasksPerExecutor).toInt
+      math.max(totalNeed, 
executorMonitor.executorCountWithResourceProfile(rpId)) +

Review comment:
       Currently `ExecutorAllocationManager` runs every 100ms and 
`resourceOffer` happens every 1sec. `unschedulableTaskSets` gets cleared only 
if any task becomes schedulable and also we'll send the `unschedulable` event 
only once unless its removed from the `unschedulableTaskSets`. So I'm guessing 
it should be fine. Please correct me if I'm forgetting something.




----------------------------------------------------------------
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.

For queries about this service, please contact Infrastructure at:
[email protected]



---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to