opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/entityownership/EntityOwnershipShard.java

   1 /*
   2  * Copyright (c) 2015 Brocade Communications Systems, Inc. and others.  All rights reserved.
   3  *
   4  * This program and the accompanying materials are made available under the
   5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   6  * and is available at http://www.eclipse.org/legal/epl-v10.html
   7  */
   8 package org.opendaylight.controller.cluster.datastore.entityownership;
   9
  10 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.CANDIDATE_NAME_NODE_ID;
  11 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.CANDIDATE_NODE_ID;
  12 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_ID_NODE_ID;
  13 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_ID_QNAME;
  14 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_NODE_ID;
  15 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_OWNERS_PATH;
  16 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_OWNER_NODE_ID;
  17 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_OWNER_QNAME;
  18 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_TYPES_PATH;
  19 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_TYPE_NODE_ID;
  20 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.ENTITY_TYPE_QNAME;
  21 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.candidateNodeKey;
  22 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.candidatePath;
  23 import static org.opendaylight.controller.cluster.datastore.entityownership.EntityOwnersModel.entityOwnersWithCandidate;
  24
  25 import akka.actor.ActorRef;
  26 import akka.actor.ActorSelection;
  27 import akka.actor.Cancellable;
  28 import akka.cluster.Cluster;
  29 import akka.cluster.ClusterEvent.CurrentClusterState;
  30 import akka.cluster.Member;
  31 import akka.cluster.MemberStatus;
  32 import akka.pattern.Patterns;
  33 import com.google.common.base.Optional;
  34 import com.google.common.base.Preconditions;
  35 import com.google.common.base.Strings;
  36 import com.google.common.collect.ImmutableSet;
  37 import java.util.ArrayList;
  38 import java.util.Collection;
  39 import java.util.HashMap;
  40 import java.util.HashSet;
  41 import java.util.Map;
  42 import java.util.Set;
  43 import java.util.concurrent.TimeUnit;
  44 import org.opendaylight.controller.cluster.access.concepts.MemberName;
  45 import org.opendaylight.controller.cluster.datastore.DatastoreContext;
  46 import org.opendaylight.controller.cluster.datastore.Shard;
  47 import org.opendaylight.controller.cluster.datastore.entityownership.messages.CandidateAdded;
  48 import org.opendaylight.controller.cluster.datastore.entityownership.messages.CandidateRemoved;
  49 import org.opendaylight.controller.cluster.datastore.entityownership.messages.RegisterCandidateLocal;
  50 import org.opendaylight.controller.cluster.datastore.entityownership.messages.RegisterListenerLocal;
  51 import org.opendaylight.controller.cluster.datastore.entityownership.messages.RemoveAllCandidates;
  52 import org.opendaylight.controller.cluster.datastore.entityownership.messages.SelectOwner;
  53 import org.opendaylight.controller.cluster.datastore.entityownership.messages.UnregisterCandidateLocal;
  54 import org.opendaylight.controller.cluster.datastore.entityownership.messages.UnregisterListenerLocal;
  55 import org.opendaylight.controller.cluster.datastore.entityownership.selectionstrategy.EntityOwnerSelectionStrategy;
  56 import org.opendaylight.controller.cluster.datastore.entityownership.selectionstrategy.EntityOwnerSelectionStrategyConfig;
  57 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  58 import org.opendaylight.controller.cluster.datastore.messages.PeerDown;
  59 import org.opendaylight.controller.cluster.datastore.messages.PeerUp;
  60 import org.opendaylight.controller.cluster.datastore.messages.SuccessReply;
  61 import org.opendaylight.controller.cluster.datastore.modification.DeleteModification;
  62 import org.opendaylight.controller.cluster.datastore.modification.MergeModification;
  63 import org.opendaylight.controller.cluster.datastore.modification.WriteModification;
  64 import org.opendaylight.controller.cluster.raft.RaftState;
  65 import org.opendaylight.mdsal.eos.dom.api.DOMEntity;
  66 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  67 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier.PathArgument;
  68 import org.opendaylight.yangtools.yang.data.api.schema.DataContainerChild;
  69 import org.opendaylight.yangtools.yang.data.api.schema.MapEntryNode;
  70 import org.opendaylight.yangtools.yang.data.api.schema.MapNode;
  71 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  72 import org.opendaylight.yangtools.yang.data.impl.schema.ImmutableNodes;
  73 import scala.concurrent.Future;
  74 import scala.concurrent.duration.FiniteDuration;
  75
  76 /**
  77  * Special Shard for EntityOwnership.
  78  *
  79  * @author Thomas Pantelis
  80  */
  81 class EntityOwnershipShard extends Shard {
  82     private final MemberName localMemberName;
  83     private final EntityOwnershipShardCommitCoordinator commitCoordinator;
  84     private final EntityOwnershipListenerSupport listenerSupport;
  85     private final Set<MemberName> downPeerMemberNames = new HashSet<>();
  86     private final EntityOwnerSelectionStrategyConfig strategyConfig;
  87     private final Map<YangInstanceIdentifier, Cancellable> entityToScheduledOwnershipTask = new HashMap<>();
  88     private final EntityOwnershipStatistics entityOwnershipStatistics;
  89     private boolean removeAllInitialCandidates = true;
  90
  91     protected EntityOwnershipShard(final Builder builder) {
  92         super(builder);
  93         this.localMemberName = builder.localMemberName;
  94         this.commitCoordinator = new EntityOwnershipShardCommitCoordinator(builder.localMemberName, LOG);
  95         this.listenerSupport = new EntityOwnershipListenerSupport(getContext(), persistenceId());
  96         this.strategyConfig = builder.ownerSelectionStrategyConfig;
  97         this.entityOwnershipStatistics = new EntityOwnershipStatistics();
  98         this.entityOwnershipStatistics.init(getDataStore());
  99     }
 100
 101     private static DatastoreContext noPersistenceDatastoreContext(final DatastoreContext datastoreContext) {
 102         return DatastoreContext.newBuilderFrom(datastoreContext).persistent(false).build();
 103     }
 104
 105     @Override
 106     protected void onDatastoreContext(final DatastoreContext context) {
 107         super.onDatastoreContext(noPersistenceDatastoreContext(context));
 108     }
 109
 110     @Override
 111     protected void onRecoveryComplete() {
 112         super.onRecoveryComplete();
 113
 114         new CandidateListChangeListener(getSelf(), persistenceId()).init(getDataStore());
 115         new EntityOwnerChangeListener(localMemberName, listenerSupport).init(getDataStore());
 116     }
 117
 118     @Override
 119     public void handleNonRaftCommand(final Object message) {
 120         if (message instanceof RegisterCandidateLocal) {
 121             onRegisterCandidateLocal((RegisterCandidateLocal) message);
 122         } else if (message instanceof UnregisterCandidateLocal) {
 123             onUnregisterCandidateLocal((UnregisterCandidateLocal) message);
 124         } else if (message instanceof CandidateAdded) {
 125             onCandidateAdded((CandidateAdded) message);
 126         } else if (message instanceof CandidateRemoved) {
 127             onCandidateRemoved((CandidateRemoved) message);
 128         } else if (message instanceof PeerDown) {
 129             onPeerDown((PeerDown) message);
 130         } else if (message instanceof PeerUp) {
 131             onPeerUp((PeerUp) message);
 132         } else if (message instanceof RegisterListenerLocal) {
 133             onRegisterListenerLocal((RegisterListenerLocal) message);
 134         } else if (message instanceof UnregisterListenerLocal) {
 135             onUnregisterListenerLocal((UnregisterListenerLocal) message);
 136         } else if (message instanceof SelectOwner) {
 137             onSelectOwner((SelectOwner) message);
 138         } else if (message instanceof RemoveAllCandidates) {
 139             onRemoveAllCandidates((RemoveAllCandidates) message);
 140         } else if (!commitCoordinator.handleMessage(message, this)) {
 141             super.handleNonRaftCommand(message);
 142         }
 143     }
 144
 145     private void onRemoveAllCandidates(final RemoveAllCandidates message) {
 146         LOG.debug("{}: onRemoveAllCandidates: {}", persistenceId(), message);
 147
 148         removeCandidateFromEntities(message.getMemberName());
 149     }
 150
 151     private void onSelectOwner(final SelectOwner selectOwner) {
 152         LOG.debug("{}: onSelectOwner: {}", persistenceId(), selectOwner);
 153
 154         String currentOwner = getCurrentOwner(selectOwner.getEntityPath());
 155         if (Strings.isNullOrEmpty(currentOwner)) {
 156             writeNewOwner(selectOwner.getEntityPath(), newOwner(currentOwner, selectOwner.getAllCandidates(),
 157                     selectOwner.getOwnerSelectionStrategy()));
 158
 159             Cancellable cancellable = entityToScheduledOwnershipTask.get(selectOwner.getEntityPath());
 160             if (cancellable != null) {
 161                 if (!cancellable.isCancelled()) {
 162                     cancellable.cancel();
 163                 }
 164                 entityToScheduledOwnershipTask.remove(selectOwner.getEntityPath());
 165             }
 166         }
 167     }
 168
 169     private void onRegisterCandidateLocal(final RegisterCandidateLocal registerCandidate) {
 170         LOG.debug("{}: onRegisterCandidateLocal: {}", persistenceId(), registerCandidate);
 171
 172         listenerSupport.setHasCandidateForEntity(registerCandidate.getEntity());
 173
 174         NormalizedNode<?, ?> entityOwners = entityOwnersWithCandidate(registerCandidate.getEntity().getType(),
 175                 registerCandidate.getEntity().getIdentifier(), localMemberName.getName());
 176         commitCoordinator.commitModification(new MergeModification(ENTITY_OWNERS_PATH, entityOwners), this);
 177
 178         getSender().tell(SuccessReply.INSTANCE, getSelf());
 179     }
 180
 181     private void onUnregisterCandidateLocal(final UnregisterCandidateLocal unregisterCandidate) {
 182         LOG.debug("{}: onUnregisterCandidateLocal: {}", persistenceId(), unregisterCandidate);
 183
 184         DOMEntity entity = unregisterCandidate.getEntity();
 185         listenerSupport.unsetHasCandidateForEntity(entity);
 186
 187         YangInstanceIdentifier candidatePath = candidatePath(entity.getType(), entity.getIdentifier(),
 188                 localMemberName.getName());
 189         commitCoordinator.commitModification(new DeleteModification(candidatePath), this);
 190
 191         getSender().tell(SuccessReply.INSTANCE, getSelf());
 192     }
 193
 194     private void onRegisterListenerLocal(final RegisterListenerLocal registerListener) {
 195         LOG.debug("{}: onRegisterListenerLocal: {}", persistenceId(), registerListener);
 196
 197         listenerSupport.addEntityOwnershipListener(registerListener.getEntityType(), registerListener.getListener());
 198
 199         getSender().tell(SuccessReply.INSTANCE, getSelf());
 200
 201         searchForEntities((entityTypeNode, entityNode) -> {
 202             Optional<DataContainerChild<?, ?>> possibleType = entityTypeNode.getChild(ENTITY_TYPE_NODE_ID);
 203             String entityType = possibleType.isPresent() ? possibleType.get().getValue().toString() : null;
 204             if (registerListener.getEntityType().equals(entityType)) {
 205                 final boolean hasOwner;
 206                 final boolean isOwner;
 207
 208                 Optional<DataContainerChild<?, ?>> possibleOwner = entityNode.getChild(ENTITY_OWNER_NODE_ID);
 209                 if (possibleOwner.isPresent()) {
 210                     isOwner = localMemberName.getName().equals(possibleOwner.get().getValue().toString());
 211                     hasOwner = true;
 212                 } else {
 213                     isOwner = false;
 214                     hasOwner = false;
 215                 }
 216
 217                 DOMEntity entity = new DOMEntity(entityType,
 218                     (YangInstanceIdentifier) entityNode.getChild(ENTITY_ID_NODE_ID).get().getValue());
 219
 220                 listenerSupport.notifyEntityOwnershipListener(entity, false, isOwner, hasOwner,
 221                     registerListener.getListener());
 222             }
 223         });
 224     }
 225
 226     private void onUnregisterListenerLocal(final UnregisterListenerLocal unregisterListener) {
 227         LOG.debug("{}: onUnregisterListenerLocal: {}", persistenceId(), unregisterListener);
 228
 229         listenerSupport.removeEntityOwnershipListener(unregisterListener.getEntityType(),
 230                 unregisterListener.getListener());
 231
 232         getSender().tell(SuccessReply.INSTANCE, getSelf());
 233     }
 234
 235     void tryCommitModifications(final BatchedModifications modifications) {
 236         if (isLeader()) {
 237             LOG.debug("{}: Committing BatchedModifications {} locally", persistenceId(),
 238                     modifications.getTransactionId());
 239
 240             // Note that it's possible the commit won't get consensus and will timeout and not be applied
 241             // to the state. However we don't need to retry it in that case b/c it will be committed to
 242             // the journal first and, once a majority of followers come back on line and it is replicated,
 243             // it will be applied at that point.
 244             handleBatchedModificationsLocal(modifications, self());
 245         } else {
 246             final ActorSelection leader = getLeader();
 247             if (leader != null) {
 248                 possiblyRemoveAllInitialCandidates(leader);
 249
 250                 LOG.debug("{}: Sending BatchedModifications {} to leader {}", persistenceId(),
 251                         modifications.getTransactionId(), leader);
 252
 253                 Future<Object> future = Patterns.ask(leader, modifications, TimeUnit.SECONDS.toMillis(
 254                         getDatastoreContext().getShardTransactionCommitTimeoutInSeconds()));
 255
 256                 Patterns.pipe(future, getContext().dispatcher()).pipeTo(getSelf(), ActorRef.noSender());
 257             }
 258         }
 259     }
 260
 261     void possiblyRemoveAllInitialCandidates(final ActorSelection leader) {
 262         // The following handles removing all candidates on startup when re-joining with a remote leader. When a
 263         // follower is detected as down, the leader will re-assign new owners to entities that were owned by the
 264         // down member but doesn't remove the down member as a candidate, as the down node may actually be isolated
 265         // and still running. Therefore on startup we send an initial message to the remote leader to remove any
 266         // potential stale candidates we had previously registered, as it's possible a candidate may not be
 267         // registered by a client in the new incarnation. We have to send the RemoveAllCandidates message prior to any
 268         // pending registrations.
 269         if (removeAllInitialCandidates && leader != null) {
 270             removeAllInitialCandidates = false;
 271             if (!isLeader()) {
 272                 LOG.debug("{} - got new leader {} on startup - sending RemoveAllCandidates", persistenceId(), leader);
 273
 274                 leader.tell(new RemoveAllCandidates(localMemberName), ActorRef.noSender());
 275             }
 276         }
 277     }
 278
 279     boolean hasLeader() {
 280         return getLeader() != null && (!isLeader() || isLeaderActive());
 281     }
 282
 283     /**
 284      * Determine if we are in jeopardy based on observed RAFT state.
 285      */
 286     private static boolean inJeopardy(final RaftState state) {
 287         switch (state) {
 288             case Candidate:
 289             case Follower:
 290             case Leader:
 291             case PreLeader:
 292                 return false;
 293             case IsolatedLeader:
 294                 return true;
 295             default:
 296                 throw new IllegalStateException("Unsupported RAFT state " + state);
 297         }
 298     }
 299
 300     private void notifyAllListeners() {
 301         searchForEntities((entityTypeNode, entityNode) -> {
 302             Optional<DataContainerChild<?, ?>> possibleType = entityTypeNode.getChild(ENTITY_TYPE_NODE_ID);
 303             if (possibleType.isPresent()) {
 304                 final boolean hasOwner;
 305                 final boolean isOwner;
 306
 307                 Optional<DataContainerChild<?, ?>> possibleOwner = entityNode.getChild(ENTITY_OWNER_NODE_ID);
 308                 if (possibleOwner.isPresent()) {
 309                     isOwner = localMemberName.getName().equals(possibleOwner.get().getValue().toString());
 310                     hasOwner = true;
 311                 } else {
 312                     isOwner = false;
 313                     hasOwner = false;
 314                 }
 315
 316                 DOMEntity entity = new DOMEntity(possibleType.get().getValue().toString(),
 317                     (YangInstanceIdentifier) entityNode.getChild(ENTITY_ID_NODE_ID).get().getValue());
 318
 319                 listenerSupport.notifyEntityOwnershipListeners(entity, isOwner, isOwner, hasOwner);
 320             }
 321         });
 322     }
 323
 324     @Override
 325     protected void onStateChanged() {
 326         boolean isLeader = isLeader();
 327         LOG.debug("{}: onStateChanged: isLeader: {}, hasLeader: {}", persistenceId(), isLeader, hasLeader());
 328
 329         // Examine current RAFT state to see if we are in jeopardy, potentially notifying all listeners
 330         final boolean inJeopardy = inJeopardy(getRaftState());
 331         final boolean wasInJeopardy = listenerSupport.setInJeopardy(inJeopardy);
 332         if (inJeopardy != wasInJeopardy) {
 333             LOG.debug("{}: {} jeopardy state, notifying all listeners", persistenceId(),
 334                 inJeopardy ? "entered" : "left");
 335             notifyAllListeners();
 336         }
 337
 338         commitCoordinator.onStateChanged(this, isLeader);
 339
 340         super.onStateChanged();
 341     }
 342
 343     @Override
 344     protected void onLeaderChanged(final String oldLeader, final String newLeader) {
 345         boolean isLeader = isLeader();
 346         LOG.debug("{}: onLeaderChanged: oldLeader: {}, newLeader: {}, isLeader: {}", persistenceId(), oldLeader,
 347                 newLeader, isLeader);
 348
 349         if (isLeader) {
 350
 351             // Re-initialize the downPeerMemberNames from the current akka Cluster state. The previous leader, if any,
 352             // is most likely down however it's possible we haven't received the PeerDown message yet.
 353             initializeDownPeerMemberNamesFromClusterState();
 354
 355             // Clear all existing strategies so that they get re-created when we call createStrategy again
 356             // This allows the strategies to be re-initialized with existing statistics maintained by
 357             // EntityOwnershipStatistics
 358             strategyConfig.clearStrategies();
 359
 360             // Re-assign owners for all members that are known to be down. In a cluster which has greater than
 361             // 3 nodes it is possible for some node beside the leader being down when the leadership transitions
 362             // it makes sense to use this event to re-assign owners for those downed nodes.
 363             Set<String> ownedBy = new HashSet<>(downPeerMemberNames.size() + 1);
 364             for (MemberName downPeerName : downPeerMemberNames) {
 365                 ownedBy.add(downPeerName.getName());
 366             }
 367
 368             // Also try to assign owners for entities that have no current owner. See explanation in onPeerUp.
 369             ownedBy.add("");
 370             selectNewOwnerForEntitiesOwnedBy(ownedBy);
 371         } else {
 372             // The leader changed - notify the coordinator to check if pending modifications need to be sent.
 373             // While onStateChanged also does this, this method handles the case where the shard hears from a
 374             // leader and stays in the follower state. In that case no behavior state change occurs.
 375             commitCoordinator.onStateChanged(this, isLeader);
 376         }
 377
 378         super.onLeaderChanged(oldLeader, newLeader);
 379     }
 380
 381     private void initializeDownPeerMemberNamesFromClusterState() {
 382         java.util.Optional<Cluster> cluster = getRaftActorContext().getCluster();
 383         if (!cluster.isPresent()) {
 384             return;
 385         }
 386
 387         CurrentClusterState state = cluster.get().state();
 388         Set<Member> unreachable = state.getUnreachable();
 389
 390         LOG.debug(
 391             "{}: initializeDownPeerMemberNamesFromClusterState - current downPeerMemberNames: {}, unreachable: {}",
 392             persistenceId(), downPeerMemberNames, unreachable);
 393
 394         downPeerMemberNames.clear();
 395         for (Member m: unreachable) {
 396             downPeerMemberNames.add(MemberName.forName(m.getRoles().iterator().next()));
 397         }
 398
 399         for (Member m: state.getMembers()) {
 400             if (m.status() != MemberStatus.up() && m.status() != MemberStatus.weaklyUp()) {
 401                 LOG.debug("{}: Adding down member with status {}", persistenceId(), m.status());
 402                 downPeerMemberNames.add(MemberName.forName(m.getRoles().iterator().next()));
 403             }
 404         }
 405
 406         LOG.debug("{}: new downPeerMemberNames: {}", persistenceId(), downPeerMemberNames);
 407     }
 408
 409     private void onCandidateRemoved(final CandidateRemoved message) {
 410         LOG.debug("{}: onCandidateRemoved: {}", persistenceId(), message);
 411
 412         if (isLeader()) {
 413             String currentOwner = getCurrentOwner(message.getEntityPath());
 414             writeNewOwner(message.getEntityPath(),
 415                     newOwner(currentOwner, message.getRemainingCandidates(),
 416                             getEntityOwnerElectionStrategy(message.getEntityPath())));
 417         }
 418     }
 419
 420     private EntityOwnerSelectionStrategy getEntityOwnerElectionStrategy(final YangInstanceIdentifier entityPath) {
 421         final String entityType = EntityOwnersModel.entityTypeFromEntityPath(entityPath);
 422         return strategyConfig.createStrategy(entityType, entityOwnershipStatistics.byEntityType(entityType));
 423     }
 424
 425     private void onCandidateAdded(final CandidateAdded message) {
 426         if (!isLeader()) {
 427             return;
 428         }
 429
 430         LOG.debug("{}: onCandidateAdded: {}", persistenceId(), message);
 431
 432         // Since a node's candidate member is only added by the node itself, we can assume the node is up so
 433         // remove it from the downPeerMemberNames.
 434         downPeerMemberNames.remove(MemberName.forName(message.getNewCandidate()));
 435
 436         final String currentOwner = getCurrentOwner(message.getEntityPath());
 437         final EntityOwnerSelectionStrategy strategy = getEntityOwnerElectionStrategy(message.getEntityPath());
 438
 439         // Available members is all the known peers - the number of peers that are down + self
 440         // So if there are 2 peers and 1 is down then availableMembers will be 2
 441         final int availableMembers = getRaftActorContext().getPeerIds().size() - downPeerMemberNames.size() + 1;
 442
 443         LOG.debug("{}: Using strategy {} to select owner, currentOwner = {}", persistenceId(), strategy, currentOwner);
 444
 445         if (strategy.getSelectionDelayInMillis() == 0L) {
 446             writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
 447                     strategy));
 448         } else if (message.getAllCandidates().size() == availableMembers) {
 449             LOG.debug("{}: Received the maximum candidates requests : {} writing new owner",
 450                     persistenceId(), availableMembers);
 451             cancelOwnerSelectionTask(message.getEntityPath());
 452             writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
 453                     strategy));
 454         } else {
 455             scheduleOwnerSelection(message.getEntityPath(), message.getAllCandidates(), strategy);
 456         }
 457     }
 458
 459     private void onPeerDown(final PeerDown peerDown) {
 460         LOG.info("{}: onPeerDown: {}", persistenceId(), peerDown);
 461
 462         MemberName downMemberName = peerDown.getMemberName();
 463         if (downPeerMemberNames.add(downMemberName) && isLeader()) {
 464             // Select new owners for entities owned by the down peer and which have other candidates. For an entity for
 465             // which the down peer is the only candidate, we leave it as the owner and don't clear it. This is done to
 466             // handle the case where the peer member process is actually still running but the node is partitioned.
 467             // When the partition is healed, the peer just remains as the owner. If the peer process actually restarted,
 468             // it will first remove all its candidates on startup. If another candidate is registered during the time
 469             // the peer is down, the new candidate will be selected as the new owner.
 470
 471             selectNewOwnerForEntitiesOwnedBy(ImmutableSet.of(downMemberName.getName()));
 472         }
 473     }
 474
 475     private void selectNewOwnerForEntitiesOwnedBy(final Set<String> ownedBy) {
 476         final BatchedModifications modifications = commitCoordinator.newBatchedModifications();
 477         searchForEntitiesOwnedBy(ownedBy, (entityTypeNode, entityNode) -> {
 478             YangInstanceIdentifier entityPath = YangInstanceIdentifier.builder(ENTITY_TYPES_PATH)
 479                     .node(entityTypeNode.getIdentifier()).node(ENTITY_NODE_ID).node(entityNode.getIdentifier())
 480                     .node(ENTITY_OWNER_NODE_ID).build();
 481             String newOwner = newOwner(getCurrentOwner(entityPath), getCandidateNames(entityNode),
 482                     getEntityOwnerElectionStrategy(entityPath));
 483
 484             if (!newOwner.isEmpty()) {
 485                 LOG.debug("{}: Found entity {}, writing new owner {}", persistenceId(), entityPath, newOwner);
 486
 487                 modifications.addModification(new WriteModification(entityPath,
 488                         ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)));
 489
 490             } else {
 491                 LOG.debug("{}: Found entity {} but no other candidates - not clearing owner", persistenceId(),
 492                         entityPath, newOwner);
 493             }
 494         });
 495
 496         commitCoordinator.commitModifications(modifications, this);
 497     }
 498
 499     private void onPeerUp(final PeerUp peerUp) {
 500         LOG.debug("{}: onPeerUp: {}", persistenceId(), peerUp);
 501
 502         downPeerMemberNames.remove(peerUp.getMemberName());
 503
 504         // Notify the coordinator to check if pending modifications need to be sent. We do this here
 505         // to handle the case where the leader's peer address isn't known yet when a prior state or
 506         // leader change occurred.
 507         commitCoordinator.onStateChanged(this, isLeader());
 508
 509         if (isLeader()) {
 510             // Try to assign owners for entities that have no current owner. It's possible the peer that is now up
 511             // had previously registered as a candidate and was the only candidate but the owner write tx couldn't be
 512             // committed due to a leader change. Eg, the leader is able to successfully commit the candidate add tx but
 513             // becomes isolated before it can commit the owner change and switches to follower. The majority partition
 514             // with a new leader has the candidate but the entity has no owner. When the partition is healed and the
 515             // previously isolated leader reconnects, we'll receive onPeerUp and, if there's still no owner, the
 516             // previous leader will gain ownership.
 517             selectNewOwnerForEntitiesOwnedBy(ImmutableSet.of(""));
 518         }
 519     }
 520
 521     private static Collection<String> getCandidateNames(final MapEntryNode entity) {
 522         Collection<MapEntryNode> candidates = ((MapNode)entity.getChild(CANDIDATE_NODE_ID).get()).getValue();
 523         Collection<String> candidateNames = new ArrayList<>(candidates.size());
 524         for (MapEntryNode candidate: candidates) {
 525             candidateNames.add(candidate.getChild(CANDIDATE_NAME_NODE_ID).get().getValue().toString());
 526         }
 527
 528         return candidateNames;
 529     }
 530
 531     private void searchForEntitiesOwnedBy(final Set<String> ownedBy, final EntityWalker walker) {
 532         LOG.debug("{}: Searching for entities owned by {}", persistenceId(), ownedBy);
 533
 534         searchForEntities((entityTypeNode, entityNode) -> {
 535             Optional<DataContainerChild<? extends PathArgument, ?>> possibleOwner =
 536                     entityNode.getChild(ENTITY_OWNER_NODE_ID);
 537             String currentOwner = possibleOwner.isPresent() ? possibleOwner.get().getValue().toString() : "";
 538             if (ownedBy.contains(currentOwner)) {
 539                 walker.onEntity(entityTypeNode, entityNode);
 540             }
 541         });
 542     }
 543
 544     private void removeCandidateFromEntities(final MemberName member) {
 545         final BatchedModifications modifications = commitCoordinator.newBatchedModifications();
 546         searchForEntities((entityTypeNode, entityNode) -> {
 547             if (hasCandidate(entityNode, member)) {
 548                 YangInstanceIdentifier entityId =
 549                         (YangInstanceIdentifier) entityNode.getIdentifier().getKeyValues().get(ENTITY_ID_QNAME);
 550                 YangInstanceIdentifier candidatePath = candidatePath(
 551                         entityTypeNode.getIdentifier().getKeyValues().get(ENTITY_TYPE_QNAME).toString(),
 552                         entityId, member.getName());
 553
 554                 LOG.info("{}: Found entity {}, removing candidate {}, path {}", persistenceId(), entityId,
 555                         member, candidatePath);
 556
 557                 modifications.addModification(new DeleteModification(candidatePath));
 558             }
 559         });
 560
 561         commitCoordinator.commitModifications(modifications, this);
 562     }
 563
 564     private static boolean hasCandidate(final MapEntryNode entity, final MemberName candidateName) {
 565         return ((MapNode)entity.getChild(CANDIDATE_NODE_ID).get()).getChild(candidateNodeKey(candidateName.getName()))
 566                 .isPresent();
 567     }
 568
 569     private void searchForEntities(final EntityWalker walker) {
 570         Optional<NormalizedNode<?, ?>> possibleEntityTypes = getDataStore().readNode(ENTITY_TYPES_PATH);
 571         if (!possibleEntityTypes.isPresent()) {
 572             return;
 573         }
 574
 575         for (MapEntryNode entityType:  ((MapNode) possibleEntityTypes.get()).getValue()) {
 576             Optional<DataContainerChild<?, ?>> possibleEntities = entityType.getChild(ENTITY_NODE_ID);
 577             if (!possibleEntities.isPresent()) {
 578                 // shouldn't happen but handle anyway
 579                 continue;
 580             }
 581
 582             for (MapEntryNode entity:  ((MapNode) possibleEntities.get()).getValue()) {
 583                 walker.onEntity(entityType, entity);
 584             }
 585         }
 586     }
 587
 588     private void writeNewOwner(final YangInstanceIdentifier entityPath, final String newOwner) {
 589         LOG.debug("{}: Writing new owner {} for entity {}", persistenceId(), newOwner, entityPath);
 590
 591         commitCoordinator.commitModification(new WriteModification(entityPath.node(ENTITY_OWNER_QNAME),
 592                 ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)), this);
 593     }
 594
 595     /**
 596      * Schedule a new owner selection job. Cancelling any outstanding job if it has not been cancelled.
 597      */
 598     private void scheduleOwnerSelection(final YangInstanceIdentifier entityPath, final Collection<String> allCandidates,
 599                                        final EntityOwnerSelectionStrategy strategy) {
 600         cancelOwnerSelectionTask(entityPath);
 601
 602         LOG.debug("{}: Scheduling owner selection after {} ms", persistenceId(), strategy.getSelectionDelayInMillis());
 603
 604         final Cancellable lastScheduledTask = context().system().scheduler().scheduleOnce(
 605                 FiniteDuration.apply(strategy.getSelectionDelayInMillis(), TimeUnit.MILLISECONDS), self(),
 606                 new SelectOwner(entityPath, allCandidates, strategy), context().system().dispatcher(), self());
 607
 608         entityToScheduledOwnershipTask.put(entityPath, lastScheduledTask);
 609     }
 610
 611     private void cancelOwnerSelectionTask(final YangInstanceIdentifier entityPath) {
 612         final Cancellable lastScheduledTask = entityToScheduledOwnershipTask.get(entityPath);
 613         if (lastScheduledTask != null && !lastScheduledTask.isCancelled()) {
 614             lastScheduledTask.cancel();
 615         }
 616     }
 617
 618     private String newOwner(final String currentOwner, final Collection<String> candidates,
 619             final EntityOwnerSelectionStrategy ownerSelectionStrategy) {
 620         Collection<String> viableCandidates = getViableCandidates(candidates);
 621         if (viableCandidates.isEmpty()) {
 622             return "";
 623         }
 624         return ownerSelectionStrategy.newOwner(currentOwner, viableCandidates);
 625     }
 626
 627     private Collection<String> getViableCandidates(final Collection<String> candidates) {
 628         Collection<String> viableCandidates = new ArrayList<>();
 629
 630         for (String candidate : candidates) {
 631             if (!downPeerMemberNames.contains(MemberName.forName(candidate))) {
 632                 viableCandidates.add(candidate);
 633             }
 634         }
 635         return viableCandidates;
 636     }
 637
 638     private String getCurrentOwner(final YangInstanceIdentifier entityId) {
 639         Optional<NormalizedNode<?, ?>> optionalEntityOwner = getDataStore().readNode(entityId.node(ENTITY_OWNER_QNAME));
 640         if (optionalEntityOwner.isPresent()) {
 641             return optionalEntityOwner.get().getValue().toString();
 642         }
 643         return null;
 644     }
 645
 646     @FunctionalInterface
 647     private interface EntityWalker {
 648         void onEntity(MapEntryNode entityTypeNode, MapEntryNode entityNode);
 649     }
 650
 651     public static Builder newBuilder() {
 652         return new Builder();
 653     }
 654
 655     static class Builder extends Shard.AbstractBuilder<Builder, EntityOwnershipShard> {
 656         private MemberName localMemberName;
 657         private EntityOwnerSelectionStrategyConfig ownerSelectionStrategyConfig;
 658
 659         protected Builder() {
 660             super(EntityOwnershipShard.class);
 661         }
 662
 663         Builder localMemberName(final MemberName newLocalMemberName) {
 664             checkSealed();
 665             this.localMemberName = newLocalMemberName;
 666             return this;
 667         }
 668
 669         Builder ownerSelectionStrategyConfig(final EntityOwnerSelectionStrategyConfig newOwnerSelectionStrategyConfig) {
 670             checkSealed();
 671             this.ownerSelectionStrategyConfig = newOwnerSelectionStrategyConfig;
 672             return this;
 673         }
 674
 675         @Override
 676         protected void verify() {
 677             super.verify();
 678             Preconditions.checkNotNull(localMemberName, "localMemberName should not be null");
 679             Preconditions.checkNotNull(ownerSelectionStrategyConfig, "ownerSelectionStrategyConfig should not be null");
 680         }
 681     }
 682 }