HBASE-25255 Master fails to initialize when creating rs group table (#2638)

Signed-off-by: Guanghao Zhang <zghao@apache.org>

hbase-server/src/main/java/org/apache/hadoop/hbase/master/HMaster.java

@@ -964,6 +964,9 @@ public class HMaster extends HRegionServer implements MasterServices {
     }
     this.assignmentManager.joinCluster();
     // The below depends on hbase:meta being online.
+    this.assignmentManager.processOfflineRegions();
+    // this must be called after the above processOfflineRegions to prevent race
+    this.assignmentManager.wakeMetaLoadedEvent();
 
     // for migrating from a version without HBASE-25099, and also for honoring the configuration
     // first.
@@ -997,7 +1000,6 @@ public class HMaster extends HRegionServer implements MasterServices {
         }
       }
     }
-    this.assignmentManager.processOfflineRegions();
     // Initialize after meta is up as below scans meta
     if (getFavoredNodesManager() != null && !maintenanceMode) {
       SnapshotOfRegionAssignmentFromMeta snapshotOfRegionAssignment =

hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/AssignmentManager.java

@@ -505,8 +505,16 @@ public class AssignmentManager {
     return metaLoadEvent.suspendIfNotReady(proc);
   }
 
+  /**
+   * This method will be called in master initialization method after calling
+   * {@link #processOfflineRegions()}, as in processOfflineRegions we will generate assign
+   * procedures for offline regions, which may be conflict with creating table.
+   * <p/>
+   * This is a bit dirty, should be reconsidered after we decide whether to keep the
+   * {@link #processOfflineRegions()} method.
+   */
   @VisibleForTesting
-  void wakeMetaLoadedEvent() {
+  public void wakeMetaLoadedEvent() {
     metaLoadEvent.wake(getProcedureScheduler());
     assert isMetaLoaded() : "expected meta to be loaded";
   }
@@ -1505,12 +1513,23 @@ public class AssignmentManager {
   // Public so can be run by the Master as part of the startup. Needs hbase:meta to be online.
   // Needs to be done after the table state manager has been started.
   public void processOfflineRegions() {
-    List<RegionInfo> offlineRegions = regionStates.getRegionStates().stream()
+    TransitRegionStateProcedure[] procs =
-      .filter(RegionState::isOffline).filter(s -> isTableEnabled(s.getRegion().getTable()))
+      regionStates.getRegionStateNodes().stream().filter(rsn -> rsn.isInState(State.OFFLINE))
-      .map(RegionState::getRegion).collect(Collectors.toList());
+        .filter(rsn -> isTableEnabled(rsn.getRegionInfo().getTable())).map(rsn -> {
-    if (!offlineRegions.isEmpty()) {
+          rsn.lock();
-      master.getMasterProcedureExecutor().submitProcedures(
+          try {
-        master.getAssignmentManager().createRoundRobinAssignProcedures(offlineRegions));
+            if (rsn.getProcedure() != null) {
+              return null;
+            } else {
+              return rsn.setProcedure(TransitRegionStateProcedure.assign(getProcedureEnvironment(),
+                rsn.getRegionInfo(), null));
+            }
+          } finally {
+            rsn.unlock();
+          }
+        }).filter(p -> p != null).toArray(TransitRegionStateProcedure[]::new);
+    if (procs.length > 0) {
+      master.getMasterProcedureExecutor().submitProcedures(procs);
     }
   }
 
@@ -1590,8 +1609,6 @@ public class AssignmentManager {
   private void loadMeta() throws IOException {
     // TODO: use a thread pool
     regionStateStore.visitMeta(new RegionMetaLoadingVisitor());
-    // every assignment is blocked until meta is loaded.
-    wakeMetaLoadedEvent();
   }
 
   /**

hbase-server/src/main/java/org/apache/hadoop/hbase/master/procedure/CreateTableProcedure.java

@@ -234,15 +234,6 @@ public class CreateTableProcedure
     }
   }
 
-  @Override
-  protected boolean waitInitialized(MasterProcedureEnv env) {
-    if (getTableName().isSystemTable()) {
-      // Creating system table is part of the initialization, so do not wait here.
-      return false;
-    }
-    return super.waitInitialized(env);
-  }
-
   private boolean prepareCreate(final MasterProcedureEnv env) throws IOException {
     final TableName tableName = getTableName();
     if (env.getMasterServices().getTableDescriptors().exists(tableName)) {

hbase-server/src/test/java/org/apache/hadoop/hbase/master/procedure/MasterProcedureTestingUtility.java

@@ -116,6 +116,7 @@ public class MasterProcedureTestingUtility {
           AssignmentManager am = env.getAssignmentManager();
           try {
             am.joinCluster();
+            am.wakeMetaLoadedEvent();
             master.setInitialized(true);
           } catch (Exception e) {
             LOG.warn("Failed to load meta", e);