YARN-11834: Capacity Scheduler Application Stuck In ACCEPTED State due to Race Condition (#7806) Contributed by Syed Shameerur Rahman.

* YARN-11834: Capacity Scheduler Application Stuck In ACCEPTED State due to Race Condition

Signed-off-by: Shilun Fan <[email protected]>

Author: shameersss1

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/AbstractLeafQueue.java

@@ -1212,6 +1212,11 @@ public CSAssignment assignContainers(Resource clusterResource,
          assignmentIterator.hasNext();) {
       FiCaSchedulerApp application = assignmentIterator.next();
 
+      // Skip processing if the application is already removed from the applicationAttemptMap
+      if (!applicationAttemptMap.containsKey(application.getApplicationAttemptId())) {
+        continue;
+      }
+
       ActivitiesLogger.APP.startAppAllocationRecording(activitiesManager,
           node, SystemClock.getInstance().getTime(), application);
 
@@ -1264,7 +1269,10 @@ public CSAssignment assignContainers(Resource clusterResource,
         userAssignable = canAssignToUser(clusterResource, application.getUser(),
             userLimit, application, candidates.getPartition(),
             currentResourceLimits);
-        if (!userAssignable && Resources.fitsIn(cul.reservation, appReserved)) {
+        if (!userAssignable
+            && Resources.fitsIn(cul.reservation, appReserved)
+            // Consider updating cul only if the application attempt is active.
+            && applicationAttemptMap.containsKey(application.getApplicationAttemptId())) {
           cul.canAssign = false;
           cul.reservation = appReserved;
         }