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