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.List;
  42 import java.util.Map;
  43 import java.util.Set;
  44 import java.util.concurrent.TimeUnit;
  45 import org.opendaylight.controller.cluster.access.concepts.MemberName;
  46 import org.opendaylight.controller.cluster.datastore.DatastoreContext;
  47 import org.opendaylight.controller.cluster.datastore.Shard;
  48 import org.opendaylight.controller.cluster.datastore.entityownership.messages.CandidateAdded;
  49 import org.opendaylight.controller.cluster.datastore.entityownership.messages.CandidateRemoved;
  50 import org.opendaylight.controller.cluster.datastore.entityownership.messages.RegisterCandidateLocal;
  51 import org.opendaylight.controller.cluster.datastore.entityownership.messages.RegisterListenerLocal;
  52 import org.opendaylight.controller.cluster.datastore.entityownership.messages.RemoveAllCandidates;
  53 import org.opendaylight.controller.cluster.datastore.entityownership.messages.SelectOwner;
  54 import org.opendaylight.controller.cluster.datastore.entityownership.messages.UnregisterCandidateLocal;
  55 import org.opendaylight.controller.cluster.datastore.entityownership.messages.UnregisterListenerLocal;
  56 import org.opendaylight.controller.cluster.datastore.entityownership.selectionstrategy.EntityOwnerSelectionStrategy;
  57 import org.opendaylight.controller.cluster.datastore.entityownership.selectionstrategy.EntityOwnerSelectionStrategyConfig;
  58 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  59 import org.opendaylight.controller.cluster.datastore.messages.PeerDown;
  60 import org.opendaylight.controller.cluster.datastore.messages.PeerUp;
  61 import org.opendaylight.controller.cluster.datastore.messages.SuccessReply;
  62 import org.opendaylight.controller.cluster.datastore.modification.DeleteModification;
  63 import org.opendaylight.controller.cluster.datastore.modification.MergeModification;
  64 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  65 import org.opendaylight.controller.cluster.datastore.modification.WriteModification;
  66 import org.opendaylight.controller.cluster.raft.RaftState;
  67 import org.opendaylight.mdsal.eos.dom.api.DOMEntity;
  68 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  69 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier.PathArgument;
  70 import org.opendaylight.yangtools.yang.data.api.schema.DataContainerChild;
  71 import org.opendaylight.yangtools.yang.data.api.schema.MapEntryNode;
  72 import org.opendaylight.yangtools.yang.data.api.schema.MapNode;
  73 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  74 import org.opendaylight.yangtools.yang.data.impl.schema.ImmutableNodes;
  75 import scala.concurrent.Future;
  76 import scala.concurrent.duration.FiniteDuration;
  77
  78 /**
  79  * Special Shard for EntityOwnership.
  80  *
  81  * @author Thomas Pantelis
  82  */
  83 class EntityOwnershipShard extends Shard {
  84     private final MemberName localMemberName;
  85     private final EntityOwnershipShardCommitCoordinator commitCoordinator;
  86     private final EntityOwnershipListenerSupport listenerSupport;
  87     private final Set<MemberName> downPeerMemberNames = new HashSet<>();
  88     private final EntityOwnerSelectionStrategyConfig strategyConfig;
  89     private final Map<YangInstanceIdentifier, Cancellable> entityToScheduledOwnershipTask = new HashMap<>();
  90     private final EntityOwnershipStatistics entityOwnershipStatistics;
  91     private boolean removeAllInitialCandidates = true;
  92
  93     protected EntityOwnershipShard(final Builder builder) {
  94         super(builder);
  95         this.localMemberName = builder.localMemberName;
  96         this.commitCoordinator = new EntityOwnershipShardCommitCoordinator(builder.localMemberName, LOG);
  97         this.listenerSupport = new EntityOwnershipListenerSupport(getContext(), persistenceId());
  98         this.strategyConfig = builder.ownerSelectionStrategyConfig;
  99         this.entityOwnershipStatistics = new EntityOwnershipStatistics();
 100         this.entityOwnershipStatistics.init(getDataStore());
 101     }
 102
 103     private static DatastoreContext noPersistenceDatastoreContext(final DatastoreContext datastoreContext) {
 104         return DatastoreContext.newBuilderFrom(datastoreContext).persistent(false).build();
 105     }
 106
 107     @Override
 108     protected void onDatastoreContext(final DatastoreContext context) {
 109         super.onDatastoreContext(noPersistenceDatastoreContext(context));
 110     }
 111
 112     @Override
 113     protected void onRecoveryComplete() {
 114         super.onRecoveryComplete();
 115
 116         new CandidateListChangeListener(getSelf(), persistenceId()).init(getDataStore());
 117         new EntityOwnerChangeListener(localMemberName, listenerSupport).init(getDataStore());
 118     }
 119
 120     @Override
 121     public void handleNonRaftCommand(final Object message) {
 122         if (message instanceof RegisterCandidateLocal) {
 123             onRegisterCandidateLocal((RegisterCandidateLocal) message);
 124         } else if (message instanceof UnregisterCandidateLocal) {
 125             onUnregisterCandidateLocal((UnregisterCandidateLocal) message);
 126         } else if (message instanceof CandidateAdded) {
 127             onCandidateAdded((CandidateAdded) message);
 128         } else if (message instanceof CandidateRemoved) {
 129             onCandidateRemoved((CandidateRemoved) message);
 130         } else if (message instanceof PeerDown) {
 131             onPeerDown((PeerDown) message);
 132         } else if (message instanceof PeerUp) {
 133             onPeerUp((PeerUp) message);
 134         } else if (message instanceof RegisterListenerLocal) {
 135             onRegisterListenerLocal((RegisterListenerLocal) message);
 136         } else if (message instanceof UnregisterListenerLocal) {
 137             onUnregisterListenerLocal((UnregisterListenerLocal) message);
 138         } else if (message instanceof SelectOwner) {
 139             onSelectOwner((SelectOwner) message);
 140         } else if (message instanceof RemoveAllCandidates) {
 141             onRemoveAllCandidates((RemoveAllCandidates) message);
 142         } else if (!commitCoordinator.handleMessage(message, this)) {
 143             super.handleNonRaftCommand(message);
 144         }
 145     }
 146
 147     private void onRemoveAllCandidates(final RemoveAllCandidates message) {
 148         LOG.debug("{}: onRemoveAllCandidates: {}", persistenceId(), message);
 149
 150         removeCandidateFromEntities(message.getMemberName());
 151     }
 152
 153     private void onSelectOwner(final SelectOwner selectOwner) {
 154         LOG.debug("{}: onSelectOwner: {}", persistenceId(), selectOwner);
 155
 156         String currentOwner = getCurrentOwner(selectOwner.getEntityPath());
 157         if (Strings.isNullOrEmpty(currentOwner)) {
 158             writeNewOwner(selectOwner.getEntityPath(), newOwner(currentOwner, selectOwner.getAllCandidates(),
 159                     selectOwner.getOwnerSelectionStrategy()));
 160
 161             Cancellable cancellable = entityToScheduledOwnershipTask.get(selectOwner.getEntityPath());
 162             if (cancellable != null) {
 163                 if (!cancellable.isCancelled()) {
 164                     cancellable.cancel();
 165                 }
 166                 entityToScheduledOwnershipTask.remove(selectOwner.getEntityPath());
 167             }
 168         }
 169     }
 170
 171     private void onRegisterCandidateLocal(final RegisterCandidateLocal registerCandidate) {
 172         LOG.debug("{}: onRegisterCandidateLocal: {}", persistenceId(), registerCandidate);
 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         YangInstanceIdentifier candidatePath = candidatePath(entity.getType(), entity.getIdentifier(),
 186                 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(final UnregisterListenerLocal unregisterListener) {
 225         LOG.debug("{}: onUnregisterListenerLocal: {}", persistenceId(), unregisterListener);
 226
 227         listenerSupport.removeEntityOwnershipListener(unregisterListener.getEntityType(),
 228                 unregisterListener.getListener());
 229
 230         getSender().tell(SuccessReply.INSTANCE, getSelf());
 231     }
 232
 233     void tryCommitModifications(final BatchedModifications modifications) {
 234         if (isLeader()) {
 235             LOG.debug("{}: Committing BatchedModifications {} locally", persistenceId(),
 236                     modifications.getTransactionId());
 237
 238             // Note that it's possible the commit won't get consensus and will timeout and not be applied
 239             // to the state. However we don't need to retry it in that case b/c it will be committed to
 240             // the journal first and, once a majority of followers come back on line and it is replicated,
 241             // it will be applied at that point.
 242             handleBatchedModificationsLocal(modifications, self());
 243         } else {
 244             final ActorSelection leader = getLeader();
 245             if (leader != null) {
 246                 possiblyRemoveAllInitialCandidates(leader);
 247
 248                 LOG.debug("{}: Sending BatchedModifications {} to leader {}", persistenceId(),
 249                         modifications.getTransactionId(), leader);
 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(final 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             default:
 294                 throw new IllegalStateException("Unsupported RAFT state " + state);
 295         }
 296     }
 297
 298     private void notifyAllListeners() {
 299         searchForEntities((entityTypeNode, entityNode) -> {
 300             Optional<DataContainerChild<?, ?>> possibleType = entityTypeNode.getChild(ENTITY_TYPE_NODE_ID);
 301             if (possibleType.isPresent()) {
 302                 final boolean hasOwner;
 303                 final boolean isOwner;
 304
 305                 Optional<DataContainerChild<?, ?>> possibleOwner = entityNode.getChild(ENTITY_OWNER_NODE_ID);
 306                 if (possibleOwner.isPresent()) {
 307                     isOwner = localMemberName.getName().equals(possibleOwner.get().getValue().toString());
 308                     hasOwner = true;
 309                 } else {
 310                     isOwner = false;
 311                     hasOwner = false;
 312                 }
 313
 314                 DOMEntity entity = new DOMEntity(possibleType.get().getValue().toString(),
 315                     (YangInstanceIdentifier) entityNode.getChild(ENTITY_ID_NODE_ID).get().getValue());
 316
 317                 listenerSupport.notifyEntityOwnershipListeners(entity, isOwner, isOwner, hasOwner);
 318             }
 319         });
 320     }
 321
 322     @Override
 323     protected void onStateChanged() {
 324         boolean isLeader = isLeader();
 325         LOG.debug("{}: onStateChanged: isLeader: {}, hasLeader: {}", persistenceId(), isLeader, hasLeader());
 326
 327         // Examine current RAFT state to see if we are in jeopardy, potentially notifying all listeners
 328         final boolean inJeopardy = inJeopardy(getRaftState());
 329         final boolean wasInJeopardy = listenerSupport.setInJeopardy(inJeopardy);
 330         if (inJeopardy != wasInJeopardy) {
 331             LOG.debug("{}: {} jeopardy state, notifying all listeners", persistenceId(),
 332                 inJeopardy ? "entered" : "left");
 333             notifyAllListeners();
 334         }
 335
 336         commitCoordinator.onStateChanged(this, isLeader);
 337
 338         super.onStateChanged();
 339     }
 340
 341     @Override
 342     protected void onLeaderChanged(final String oldLeader, final String newLeader) {
 343         boolean isLeader = isLeader();
 344         LOG.debug("{}: onLeaderChanged: oldLeader: {}, newLeader: {}, isLeader: {}", persistenceId(), oldLeader,
 345                 newLeader, isLeader);
 346
 347         if (isLeader) {
 348
 349             // Re-initialize the downPeerMemberNames from the current akka Cluster state. The previous leader, if any,
 350             // is most likely down however it's possible we haven't received the PeerDown message yet.
 351             initializeDownPeerMemberNamesFromClusterState();
 352
 353             // Clear all existing strategies so that they get re-created when we call createStrategy again
 354             // This allows the strategies to be re-initialized with existing statistics maintained by
 355             // EntityOwnershipStatistics
 356             strategyConfig.clearStrategies();
 357
 358             // Re-assign owners for all members that are known to be down. In a cluster which has greater than
 359             // 3 nodes it is possible for some node beside the leader being down when the leadership transitions
 360             // it makes sense to use this event to re-assign owners for those downed nodes.
 361             Set<String> ownedBy = new HashSet<>(downPeerMemberNames.size() + 1);
 362             for (MemberName downPeerName : downPeerMemberNames) {
 363                 ownedBy.add(downPeerName.getName());
 364             }
 365
 366             // Also try to assign owners for entities that have no current owner. See explanation in onPeerUp.
 367             ownedBy.add("");
 368             selectNewOwnerForEntitiesOwnedBy(ownedBy);
 369         } else {
 370             // The leader changed - notify the coordinator to check if pending modifications need to be sent.
 371             // While onStateChanged also does this, this method handles the case where the shard hears from a
 372             // leader and stays in the follower state. In that case no behavior state change occurs.
 373             commitCoordinator.onStateChanged(this, isLeader);
 374         }
 375
 376         super.onLeaderChanged(oldLeader, newLeader);
 377     }
 378
 379     private void initializeDownPeerMemberNamesFromClusterState() {
 380         java.util.Optional<Cluster> cluster = getRaftActorContext().getCluster();
 381         if (!cluster.isPresent()) {
 382             return;
 383         }
 384
 385         CurrentClusterState state = cluster.get().state();
 386         Set<Member> unreachable = state.getUnreachable();
 387
 388         LOG.debug(
 389             "{}: initializeDownPeerMemberNamesFromClusterState - current downPeerMemberNames: {}, unreachable: {}",
 390             persistenceId(), downPeerMemberNames, unreachable);
 391
 392         downPeerMemberNames.clear();
 393         for (Member m: unreachable) {
 394             downPeerMemberNames.add(MemberName.forName(m.getRoles().iterator().next()));
 395         }
 396
 397         for (Member m: state.getMembers()) {
 398             if (m.status() != MemberStatus.up() && m.status() != MemberStatus.weaklyUp()) {
 399                 LOG.debug("{}: Adding down member with status {}", persistenceId(), m.status());
 400                 downPeerMemberNames.add(MemberName.forName(m.getRoles().iterator().next()));
 401             }
 402         }
 403
 404         LOG.debug("{}: new downPeerMemberNames: {}", persistenceId(), downPeerMemberNames);
 405     }
 406
 407     private void onCandidateRemoved(final CandidateRemoved message) {
 408         LOG.debug("{}: onCandidateRemoved: {}", persistenceId(), message);
 409
 410         if (isLeader()) {
 411             String currentOwner = getCurrentOwner(message.getEntityPath());
 412             writeNewOwner(message.getEntityPath(),
 413                     newOwner(currentOwner, message.getRemainingCandidates(),
 414                             getEntityOwnerElectionStrategy(message.getEntityPath())));
 415         }
 416     }
 417
 418     private EntityOwnerSelectionStrategy getEntityOwnerElectionStrategy(final YangInstanceIdentifier entityPath) {
 419         final String entityType = EntityOwnersModel.entityTypeFromEntityPath(entityPath);
 420         return strategyConfig.createStrategy(entityType, entityOwnershipStatistics.byEntityType(entityType));
 421     }
 422
 423     private void onCandidateAdded(final CandidateAdded message) {
 424         if (!isLeader()) {
 425             return;
 426         }
 427
 428         LOG.debug("{}: onCandidateAdded: {}", persistenceId(), message);
 429
 430         // Since a node's candidate member is only added by the node itself, we can assume the node is up so
 431         // remove it from the downPeerMemberNames.
 432         downPeerMemberNames.remove(MemberName.forName(message.getNewCandidate()));
 433
 434         final String currentOwner = getCurrentOwner(message.getEntityPath());
 435         final EntityOwnerSelectionStrategy strategy = getEntityOwnerElectionStrategy(message.getEntityPath());
 436
 437         // Available members is all the known peers - the number of peers that are down + self
 438         // So if there are 2 peers and 1 is down then availableMembers will be 2
 439         final int availableMembers = getRaftActorContext().getPeerIds().size() - downPeerMemberNames.size() + 1;
 440
 441         LOG.debug("{}: Using strategy {} to select owner, currentOwner = {}", persistenceId(), strategy, currentOwner);
 442
 443         if (strategy.getSelectionDelayInMillis() == 0L) {
 444             writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
 445                     strategy));
 446         } else if (message.getAllCandidates().size() == availableMembers) {
 447             LOG.debug("{}: Received the maximum candidates requests : {} writing new owner",
 448                     persistenceId(), availableMembers);
 449             cancelOwnerSelectionTask(message.getEntityPath());
 450             writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
 451                     strategy));
 452         } else {
 453             scheduleOwnerSelection(message.getEntityPath(), message.getAllCandidates(), strategy);
 454         }
 455     }
 456
 457     private void onPeerDown(final PeerDown peerDown) {
 458         LOG.info("{}: onPeerDown: {}", persistenceId(), peerDown);
 459
 460         MemberName downMemberName = peerDown.getMemberName();
 461         if (downPeerMemberNames.add(downMemberName) && isLeader()) {
 462             // Select new owners for entities owned by the down peer and which have other candidates. For an entity for
 463             // which the down peer is the only candidate, we leave it as the owner and don't clear it. This is done to
 464             // handle the case where the peer member process is actually still running but the node is partitioned.
 465             // When the partition is healed, the peer just remains as the owner. If the peer process actually restarted,
 466             // it will first remove all its candidates on startup. If another candidate is registered during the time
 467             // the peer is down, the new candidate will be selected as the new owner.
 468
 469             selectNewOwnerForEntitiesOwnedBy(ImmutableSet.of(downMemberName.getName()));
 470         }
 471     }
 472
 473     private void selectNewOwnerForEntitiesOwnedBy(final Set<String> ownedBy) {
 474         final List<Modification> modifications = new ArrayList<>();
 475         searchForEntitiesOwnedBy(ownedBy, (entityTypeNode, entityNode) -> {
 476             YangInstanceIdentifier entityPath = YangInstanceIdentifier.builder(ENTITY_TYPES_PATH)
 477                     .node(entityTypeNode.getIdentifier()).node(ENTITY_NODE_ID).node(entityNode.getIdentifier())
 478                     .node(ENTITY_OWNER_NODE_ID).build();
 479             String newOwner = newOwner(getCurrentOwner(entityPath), getCandidateNames(entityNode),
 480                     getEntityOwnerElectionStrategy(entityPath));
 481
 482             if (!newOwner.isEmpty()) {
 483                 LOG.debug("{}: Found entity {}, writing new owner {}", persistenceId(), entityPath, newOwner);
 484
 485                 modifications.add(new WriteModification(entityPath,
 486                     ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)));
 487
 488             } else {
 489                 LOG.debug("{}: Found entity {} but no other candidates - not clearing owner", persistenceId(),
 490                         entityPath, newOwner);
 491             }
 492         });
 493
 494         commitCoordinator.commitModifications(modifications, this);
 495     }
 496
 497     private void onPeerUp(final PeerUp peerUp) {
 498         LOG.debug("{}: onPeerUp: {}", persistenceId(), peerUp);
 499
 500         downPeerMemberNames.remove(peerUp.getMemberName());
 501
 502         // Notify the coordinator to check if pending modifications need to be sent. We do this here
 503         // to handle the case where the leader's peer address isn't known yet when a prior state or
 504         // leader change occurred.
 505         commitCoordinator.onStateChanged(this, isLeader());
 506
 507         if (isLeader()) {
 508             // Try to assign owners for entities that have no current owner. It's possible the peer that is now up
 509             // had previously registered as a candidate and was the only candidate but the owner write tx couldn't be
 510             // committed due to a leader change. Eg, the leader is able to successfully commit the candidate add tx but
 511             // becomes isolated before it can commit the owner change and switches to follower. The majority partition
 512             // with a new leader has the candidate but the entity has no owner. When the partition is healed and the
 513             // previously isolated leader reconnects, we'll receive onPeerUp and, if there's still no owner, the
 514             // previous leader will gain ownership.
 515             selectNewOwnerForEntitiesOwnedBy(ImmutableSet.of(""));
 516         }
 517     }
 518
 519     private static Collection<String> getCandidateNames(final MapEntryNode entity) {
 520         Collection<MapEntryNode> candidates = ((MapNode)entity.getChild(CANDIDATE_NODE_ID).get()).getValue();
 521         Collection<String> candidateNames = new ArrayList<>(candidates.size());
 522         for (MapEntryNode candidate: candidates) {
 523             candidateNames.add(candidate.getChild(CANDIDATE_NAME_NODE_ID).get().getValue().toString());
 524         }
 525
 526         return candidateNames;
 527     }
 528
 529     private void searchForEntitiesOwnedBy(final Set<String> ownedBy, final EntityWalker walker) {
 530         LOG.debug("{}: Searching for entities owned by {}", persistenceId(), ownedBy);
 531
 532         searchForEntities((entityTypeNode, entityNode) -> {
 533             Optional<DataContainerChild<? extends PathArgument, ?>> possibleOwner =
 534                     entityNode.getChild(ENTITY_OWNER_NODE_ID);
 535             String currentOwner = possibleOwner.isPresent() ? possibleOwner.get().getValue().toString() : "";
 536             if (ownedBy.contains(currentOwner)) {
 537                 walker.onEntity(entityTypeNode, entityNode);
 538             }
 539         });
 540     }
 541
 542     private void removeCandidateFromEntities(final MemberName member) {
 543         final List<Modification> modifications = new ArrayList<>();
 544         searchForEntities((entityTypeNode, entityNode) -> {
 545             if (hasCandidate(entityNode, member)) {
 546                 YangInstanceIdentifier entityId =
 547                         (YangInstanceIdentifier) entityNode.getIdentifier().getKeyValues().get(ENTITY_ID_QNAME);
 548                 YangInstanceIdentifier candidatePath = candidatePath(
 549                         entityTypeNode.getIdentifier().getKeyValues().get(ENTITY_TYPE_QNAME).toString(),
 550                         entityId, member.getName());
 551
 552                 LOG.info("{}: Found entity {}, removing candidate {}, path {}", persistenceId(), entityId,
 553                         member, candidatePath);
 554
 555                 modifications.add(new DeleteModification(candidatePath));
 556             }
 557         });
 558
 559         commitCoordinator.commitModifications(modifications, this);
 560     }
 561
 562     private static boolean hasCandidate(final MapEntryNode entity, final MemberName candidateName) {
 563         return ((MapNode)entity.getChild(CANDIDATE_NODE_ID).get()).getChild(candidateNodeKey(candidateName.getName()))
 564                 .isPresent();
 565     }
 566
 567     private void searchForEntities(final EntityWalker walker) {
 568         Optional<NormalizedNode<?, ?>> possibleEntityTypes = getDataStore().readNode(ENTITY_TYPES_PATH);
 569         if (!possibleEntityTypes.isPresent()) {
 570             return;
 571         }
 572
 573         for (MapEntryNode entityType:  ((MapNode) possibleEntityTypes.get()).getValue()) {
 574             Optional<DataContainerChild<?, ?>> possibleEntities = entityType.getChild(ENTITY_NODE_ID);
 575             if (!possibleEntities.isPresent()) {
 576                 // shouldn't happen but handle anyway
 577                 continue;
 578             }
 579
 580             for (MapEntryNode entity:  ((MapNode) possibleEntities.get()).getValue()) {
 581                 walker.onEntity(entityType, entity);
 582             }
 583         }
 584     }
 585
 586     private void writeNewOwner(final YangInstanceIdentifier entityPath, final String newOwner) {
 587         LOG.debug("{}: Writing new owner {} for entity {}", persistenceId(), newOwner, entityPath);
 588
 589         commitCoordinator.commitModification(new WriteModification(entityPath.node(ENTITY_OWNER_QNAME),
 590                 ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)), this);
 591     }
 592
 593     /**
 594      * Schedule a new owner selection job. Cancelling any outstanding job if it has not been cancelled.
 595      */
 596     private void scheduleOwnerSelection(final YangInstanceIdentifier entityPath, final Collection<String> allCandidates,
 597                                        final 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), self(),
 604                 new SelectOwner(entityPath, allCandidates, strategy), context().system().dispatcher(), self());
 605
 606         entityToScheduledOwnershipTask.put(entityPath, lastScheduledTask);
 607     }
 608
 609     private void cancelOwnerSelectionTask(final YangInstanceIdentifier entityPath) {
 610         final Cancellable lastScheduledTask = entityToScheduledOwnershipTask.get(entityPath);
 611         if (lastScheduledTask != null && !lastScheduledTask.isCancelled()) {
 612             lastScheduledTask.cancel();
 613         }
 614     }
 615
 616     private String newOwner(final String currentOwner, final Collection<String> candidates,
 617             final 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(final 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(final 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(final MemberName newLocalMemberName) {
 662             checkSealed();
 663             this.localMemberName = newLocalMemberName;
 664             return this;
 665         }
 666
 667         Builder ownerSelectionStrategyConfig(final EntityOwnerSelectionStrategyConfig newOwnerSelectionStrategyConfig) {
 668             checkSealed();
 669             this.ownerSelectionStrategyConfig = newOwnerSelectionStrategyConfig;
 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 }