Add a condition in RMAppManager to force application queue placement based on ACL during upgrade (apache#6)

This commit is only useful to perform the upgrade from 2 to 3. Once the upgrade is finalized, this specific code can (and should) be removed. See below for detailed explanation.

In HDP2, when an application was assigned to a queue through acl queue mappings, it was done at the CapacityScheduler level and not modifying the ApplicationSubmissionContext. Because of this, the application was serialized in the ZKStateStore with the queue it was submitted on at submit time, which is 'default' (we ask user not to decalre the queue but to let acl queue mappings do the job).
It was not a problem for the RM v2, because even if it read the queue default, it would eventually place it on the correct queue.

However, in HDP3, this mechanism has changed and has been rationalized through specific classes and no more the CapacityScheduler (code is a lot cleaner by the way). As a result, queue placement is evaluated before the application is completly submitted. Thus, RM v3 will serialize the final queue to which the application was assigned. But it means that with v3, at restart, the RM will read the queue field that was serialized in the state store and try to place the queue on it. It will not reevaluate the acl queue mapping.

When we upgrade from v2 to v3, the RM will read a lot of application placed on queue default and not try to perform acl queue mapping. Since there is no queue default, teh RM will KILL all the applications that were running.

The fix consists in detecting that the queue returned from the state store is 'default' and if so, evaluate the placement of the application, allowing migration without killing existing apps. Once all applciations have been started/serialized again (through app state changes) in the ZkStateStore, this code has no value anymore.

Co-authored-by: William Montaz <[email protected]>

Author: Willymontaz

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/RMAppManager.java

@@ -84,7 +84,7 @@
 /**
  * This class manages the list of applications for the resource manager. 
  */
-public class RMAppManager implements EventHandler<RMAppManagerEvent>, 
+public class RMAppManager implements EventHandler<RMAppManagerEvent>,
                                         Recoverable {
 
   private static final Logger LOG =
@@ -399,13 +399,13 @@ private RMAppImpl createAndPopulateNewRMApp(
       RMAppState recoveredFinalState) throws YarnException {
 
     ApplicationPlacementContext placementContext = null;
-    if (recoveredFinalState == null) {
+    if (recoveredFinalState == null || (isRecovery && submissionContext.getQueue().equals("default"))) {
       placementContext = placeApplication(rmContext.getQueuePlacementManager(),
           submissionContext, user, isRecovery);
     }
 
     // We only replace the queue when it's a new application
-    if (!isRecovery) {
+    if (!isRecovery || submissionContext.getQueue().equals("default")) {
       copyPlacementQueueToSubmissionContext(placementContext,
           submissionContext);