[GitHub] [spark] holdenk commented on a change in pull request #33492: [SPARK-36052][K8S] Introducing a limit for pending PODs

[GitBox <gi...@apache.org>]

holdenk commented on a change in pull request #33492:
URL: https://github.com/apache/spark/pull/33492#discussion_r675906423



##########
File path: resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodsAllocator.scala
##########
@@ -293,32 +302,49 @@ private[spark] class ExecutorPodsAllocator(
               .withLabelIn(SPARK_EXECUTOR_ID_LABEL, toDelete.sorted.map(_.toString): _*)
               .delete()
             newlyCreatedExecutors --= newlyCreatedToDelete
-            knownPendingCount -= knownPendingToDelete.size
+            pendingCountForRpId -= pendingToDelete.size
+            notRunningPodCountForRpId -= toDelete.size
           }
         }
       }
-
-      if (newlyCreatedExecutorsForRpId.isEmpty
-        && knownPodCount < targetNum) {
-        requestNewExecutors(targetNum, knownPodCount, applicationId, rpId, k8sKnownPVCNames)
-      }
-      totalPendingCount += knownPendingCount
+      totalPendingCount += pendingCountForRpId
+      totalNotRunningPodCount += notRunningPodCountForRpId
 
       // The code below just prints debug messages, which are only useful when there's a change
       // in the snapshot state. Since the messages are a little spammy, avoid them when we know
       // there are no useful updates.
       if (log.isDebugEnabled && snapshots.nonEmpty) {
-        val outstanding = knownPendingCount + newlyCreatedExecutorsForRpId.size
+        val outstanding = pendingCountForRpId + newlyCreatedExecutorsForRpId.size
         if (currentRunningCount >= targetNum && !dynamicAllocationEnabled) {
           logDebug(s"Current number of running executors for ResourceProfile Id $rpId is " +
             "equal to the number of requested executors. Not scaling up further.")
         } else {
-          if (outstanding > 0) {
-            logDebug(s"Still waiting for $outstanding executors for ResourceProfile " +
-              s"Id $rpId before requesting more.")
+          if (newlyCreatedExecutorsForRpId.nonEmpty) {
+            logDebug(s"Still waiting for ${newlyCreatedExecutorsForRpId.size} executors for " +
+              s"ResourceProfile Id $rpId before requesting more.")
           }
         }
       }
+      if (newlyCreatedExecutorsForRpId.isEmpty && podCountForRpId < targetNum) {
+        Some(rpId, podCountForRpId, targetNum)
+      } else {
+        // for this resource profile we do not request more PODs
+        None
+      }
+    }

Review comment:
       I get this is the old behaviour, but I would think that we might allocate up to max pending so requiring newlyCreated to be empty seems strange to me.

##########
File path: resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/Config.scala
##########
@@ -568,6 +568,16 @@ private[spark] object Config extends Logging {
       .checkValue(delay => delay > 0, "delay must be a positive time value")
       .createWithDefaultString("30s")
 
+  val KUBERNETES_MAX_PENDING_PODS =
+    ConfigBuilder("spark.kubernetes.allocation.maxPendingPods")
+      .doc("Maximum number of pending PODs allowed during executor allocation for this " +
+        "application. Those newly requested executors which are unknown by Kubernetes yet are " +
+        "also counted into this limit as they will change into pending PODs by time.")

Review comment:
       Maybe include a note that if there are pending pods for a given resource profile we won't allocate more pods for the rpi since on my first reading of this I assumed this flag changed that behaviour.




-- 
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: reviews-unsubscribe@spark.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org



---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org