+ private void onRegisterListenerLocal(final RegisterListenerLocal registerListener) {
+ LOG.debug("{}: onRegisterListenerLocal: {}", persistenceId(), registerListener);
+
+ listenerSupport.addEntityOwnershipListener(registerListener.getEntityType(), registerListener.getListener());
+
+ getSender().tell(SuccessReply.INSTANCE, getSelf());
+
+ searchForEntities((entityTypeNode, entityNode) -> {
+ Optional<DataContainerChild<?, ?>> possibleType = entityTypeNode.getChild(ENTITY_TYPE_NODE_ID);
+ String entityType = possibleType.isPresent() ? possibleType.get().getValue().toString() : null;
+ if (registerListener.getEntityType().equals(entityType)) {
+ final boolean hasOwner;
+ final boolean isOwner;
+
+ Optional<DataContainerChild<?, ?>> possibleOwner = entityNode.getChild(ENTITY_OWNER_NODE_ID);
+ if (possibleOwner.isPresent()) {
+ isOwner = localMemberName.getName().equals(possibleOwner.get().getValue().toString());
+ hasOwner = true;
+ } else {
+ isOwner = false;
+ hasOwner = false;
+ }
+
+ DOMEntity entity = new DOMEntity(entityType,
+ (YangInstanceIdentifier) entityNode.getChild(ENTITY_ID_NODE_ID).get().getValue());
+
+ listenerSupport.notifyEntityOwnershipListener(entity, false, isOwner, hasOwner,
+ registerListener.getListener());
+ }
+ });
+ }
+
+ private void onUnregisterListenerLocal(UnregisterListenerLocal unregisterListener) {
+ LOG.debug("{}: onUnregisterListenerLocal: {}", persistenceId(), unregisterListener);
+
+ listenerSupport.removeEntityOwnershipListener(unregisterListener.getEntityType(), unregisterListener.getListener());
+
+ getSender().tell(SuccessReply.INSTANCE, getSelf());
+ }
+
+ void tryCommitModifications(final BatchedModifications modifications) {
+ if(isLeader()) {
+ LOG.debug("{}: Committing BatchedModifications {} locally", persistenceId(), modifications.getTransactionID());
+
+ // Note that it's possible the commit won't get consensus and will timeout and not be applied
+ // to the state. However we don't need to retry it in that case b/c it will be committed to
+ // the journal first and, once a majority of followers come back on line and it is replicated,
+ // it will be applied at that point.
+ handleBatchedModificationsLocal(modifications, self());
+ } else {
+ final ActorSelection leader = getLeader();
+ if (leader != null) {
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Sending BatchedModifications {} to leader {}", persistenceId(),
+ modifications.getTransactionID(), leader);
+ }
+
+ Future<Object> future = Patterns.ask(leader, modifications, TimeUnit.SECONDS.toMillis(
+ getDatastoreContext().getShardTransactionCommitTimeoutInSeconds()));
+
+ Patterns.pipe(future, getContext().dispatcher()).pipeTo(getSelf(), ActorRef.noSender());
+ }
+ }
+ }
+
+ boolean hasLeader() {
+ return getLeader() != null && !isIsolatedLeader();
+ }
+
+ /**
+ * Determine if we are in jeopardy based on observed RAFT state.
+ */
+ private static boolean inJeopardy(final RaftState state) {
+ switch (state) {
+ case Candidate:
+ case Follower:
+ case Leader:
+ return false;
+ case IsolatedLeader:
+ return true;
+ }
+ throw new IllegalStateException("Unsupported RAFT state " + state);
+ }
+
+ private void notifyAllListeners() {
+ searchForEntities((entityTypeNode, entityNode) -> {
+ Optional<DataContainerChild<?, ?>> possibleType = entityTypeNode.getChild(ENTITY_TYPE_NODE_ID);
+ if (possibleType.isPresent()) {
+ final boolean hasOwner;
+ final boolean isOwner;
+
+ Optional<DataContainerChild<?, ?>> possibleOwner = entityNode.getChild(ENTITY_OWNER_NODE_ID);
+ if (possibleOwner.isPresent()) {
+ isOwner = localMemberName.getName().equals(possibleOwner.get().getValue().toString());
+ hasOwner = true;
+ } else {
+ isOwner = false;
+ hasOwner = false;
+ }
+
+ DOMEntity entity = new DOMEntity(possibleType.get().getValue().toString(),
+ (YangInstanceIdentifier) entityNode.getChild(ENTITY_ID_NODE_ID).get().getValue());
+
+ listenerSupport.notifyEntityOwnershipListeners(entity, isOwner, isOwner, hasOwner);
+ }
+ });
+ }
+
+ @Override
+ protected void onStateChanged() {
+ super.onStateChanged();
+
+ boolean isLeader = isLeader();
+ LOG.debug("{}: onStateChanged: isLeader: {}, hasLeader: {}", persistenceId(), isLeader, hasLeader());
+
+ // Examine current RAFT state to see if we are in jeopardy, potentially notifying all listeners
+ final boolean inJeopardy = inJeopardy(getRaftState());
+ final boolean wasInJeopardy = listenerSupport.setInJeopardy(inJeopardy);
+ if (inJeopardy != wasInJeopardy) {
+ LOG.debug("{}: {} jeopardy state, notifying all listeners", persistenceId(),
+ inJeopardy ? "entered" : "left");
+ notifyAllListeners();
+ }
+
+ commitCoordinator.onStateChanged(this, isLeader);
+ }
+
+ @Override
+ protected void onLeaderChanged(String oldLeader, String newLeader) {
+ super.onLeaderChanged(oldLeader, newLeader);
+
+ boolean isLeader = isLeader();
+ LOG.debug("{}: onLeaderChanged: oldLeader: {}, newLeader: {}, isLeader: {}", persistenceId(), oldLeader,
+ newLeader, isLeader);
+
+ if (isLeader) {
+
+ // Clear all existing strategies so that they get re-created when we call createStrategy again
+ // This allows the strategies to be re-initialized with existing statistics maintained by
+ // EntityOwnershipStatistics
+ strategyConfig.clearStrategies();
+
+ // Remove the candidates for all members that are known to be down. In a cluster which has greater than
+ // 3 nodes it is possible for a some node beside the leader being down when the leadership transitions
+ // it makes sense to use this event to remove all the candidates for those downed nodes
+ for (MemberName downPeerName : downPeerMemberNames) {
+ removeCandidateFromEntities(downPeerName);
+ }
+ } else {
+ // The leader changed - notify the coordinator to check if pending modifications need to be sent.
+ // While onStateChanged also does this, this method handles the case where the shard hears from a
+ // leader and stays in the follower state. In that case no behavior state change occurs.
+ commitCoordinator.onStateChanged(this, isLeader);
+ }
+ }
+
+ private void onCandidateRemoved(CandidateRemoved message) {
+ LOG.debug("{}: onCandidateRemoved: {}", persistenceId(), message);
+
+ if(isLeader()) {
+ String currentOwner = getCurrentOwner(message.getEntityPath());
+ if(message.getRemovedCandidate().equals(currentOwner) || message.getRemainingCandidates().size() == 0){
+ String entityType = EntityOwnersModel.entityTypeFromEntityPath(message.getEntityPath());
+ writeNewOwner(message.getEntityPath(),
+ newOwner(currentOwner, message.getRemainingCandidates(), entityOwnershipStatistics.byEntityType(entityType),
+ getEntityOwnerElectionStrategy(message.getEntityPath())));
+ }
+ } else {
+ // We're not the leader. If the removed candidate is our local member then check if we actually
+ // have a local candidate registered. If we do then we must have been partitioned from the leader
+ // and the leader removed our candidate since the leader can't tell the difference between a
+ // temporary network partition and a node's process actually restarted. So, in that case, re-add
+ // our candidate.
+ if(localMemberName.getName().equals(message.getRemovedCandidate()) &&
+ listenerSupport.hasCandidateForEntity(createEntity(message.getEntityPath()))) {
+ LOG.debug("Local candidate member was removed but a local candidate is registered for {}" +
+ " - adding back local candidate", message.getEntityPath());
+
+ commitCoordinator.commitModification(new MergeModification(
+ candidatePath(message.getEntityPath(), localMemberName.getName()),
+ candidateMapEntry(localMemberName.getName())), this);
+ }
+ }
+ }
+
+ private EntityOwnerSelectionStrategy getEntityOwnerElectionStrategy(YangInstanceIdentifier entityPath) {
+ final String entityType = EntityOwnersModel.entityTypeFromEntityPath(entityPath);
+ return strategyConfig.createStrategy(entityType, entityOwnershipStatistics.byEntityType(entityType));
+ }
+
+ private void onCandidateAdded(CandidateAdded message) {
+ if(!isLeader()){
+ return;
+ }
+
+ LOG.debug("{}: onCandidateAdded: {}", persistenceId(), message);
+
+ // Since a node's candidate member is only added by the node itself, we can assume the node is up so
+ // remove it from the downPeerMemberNames.
+ downPeerMemberNames.remove(message.getNewCandidate());
+
+ final String currentOwner = getCurrentOwner(message.getEntityPath());
+ final EntityOwnerSelectionStrategy strategy = getEntityOwnerElectionStrategy(message.getEntityPath());
+ final String entityType = EntityOwnersModel.entityTypeFromEntityPath(message.getEntityPath());
+
+ // Available members is all the known peers - the number of peers that are down + self
+ // So if there are 2 peers and 1 is down then availableMembers will be 2
+ final int availableMembers = (peerIdToMemberNames.size() - downPeerMemberNames.size()) + 1;
+
+ LOG.debug("{}: Using strategy {} to select owner, currentOwner = {}", persistenceId(), strategy, currentOwner);
+
+ if(!message.getAllCandidates().contains(currentOwner)){
+ if(strategy.getSelectionDelayInMillis() == 0L) {
+ writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
+ entityOwnershipStatistics.byEntityType(entityType), strategy));
+ } else if(message.getAllCandidates().size() == availableMembers) {
+ LOG.debug("{}: Received the maximum candidates requests : {} writing new owner",
+ persistenceId(), availableMembers);
+ cancelOwnerSelectionTask(message.getEntityPath());
+ writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
+ entityOwnershipStatistics.byEntityType(entityType), strategy));
+ } else {
+ scheduleOwnerSelection(message.getEntityPath(), message.getAllCandidates(), strategy);
+ }
+ }
+ }
+
+ private void onPeerDown(PeerDown peerDown) {
+ LOG.info("{}: onPeerDown: {}", persistenceId(), peerDown);
+
+ MemberName downMemberName = peerDown.getMemberName();
+ if(downPeerMemberNames.add(downMemberName) && isLeader()) {
+ // Remove the down peer as a candidate from all entities.
+ removeCandidateFromEntities(downMemberName);
+ }
+ }
+
+ private void onPeerUp(PeerUp peerUp) {
+ LOG.debug("{}: onPeerUp: {}", persistenceId(), peerUp);
+
+ peerIdToMemberNames.put(peerUp.getPeerId(), peerUp.getMemberName());
+ downPeerMemberNames.remove(peerUp.getMemberName());
+
+ // Notify the coordinator to check if pending modifications need to be sent. We do this here
+ // to handle the case where the leader's peer address isn't now yet when a prior state or
+ // leader change occurred.
+ commitCoordinator.onStateChanged(this, isLeader());
+ }
+
+ private void removeCandidateFromEntities(final MemberName owner) {
+ final BatchedModifications modifications = commitCoordinator.newBatchedModifications();
+ searchForEntities((entityTypeNode, entityNode) -> {
+ if (hasCandidate(entityNode, owner)) {
+ YangInstanceIdentifier entityId =
+ (YangInstanceIdentifier) entityNode.getIdentifier().getKeyValues().get(ENTITY_ID_QNAME);
+ YangInstanceIdentifier candidatePath = candidatePath(
+ entityTypeNode.getIdentifier().getKeyValues().get(ENTITY_TYPE_QNAME).toString(),
+ entityId, owner.getName());
+
+ LOG.info("{}: Found entity {}, removing candidate {}, path {}", persistenceId(), entityId,
+ owner, candidatePath);
+
+ modifications.addModification(new DeleteModification(candidatePath));
+ }
+ });
+
+ commitCoordinator.commitModifications(modifications, this);
+ }
+
+ private static boolean hasCandidate(MapEntryNode entity, MemberName candidateName) {
+ return ((MapNode)entity.getChild(CANDIDATE_NODE_ID).get()).getChild(candidateNodeKey(candidateName.getName()))
+ .isPresent();
+ }
+
+ private void searchForEntities(EntityWalker walker) {
+ Optional<NormalizedNode<?, ?>> possibleEntityTypes = getDataStore().readNode(ENTITY_TYPES_PATH);
+ if(!possibleEntityTypes.isPresent()) {
+ return;
+ }
+
+ for(MapEntryNode entityType: ((MapNode) possibleEntityTypes.get()).getValue()) {
+ Optional<DataContainerChild<?, ?>> possibleEntities = entityType.getChild(ENTITY_NODE_ID);
+ if(!possibleEntities.isPresent()) {
+ // shouldn't happen but handle anyway
+ continue;
+ }
+
+ for(MapEntryNode entity: ((MapNode) possibleEntities.get()).getValue()) {
+ walker.onEntity(entityType, entity);
+ }
+ }
+ }
+
+ private void writeNewOwner(YangInstanceIdentifier entityPath, String newOwner) {
+ LOG.debug("{}: Writing new owner {} for entity {}", persistenceId(), newOwner, entityPath);
+
+ commitCoordinator.commitModification(new WriteModification(entityPath.node(ENTITY_OWNER_QNAME),
+ ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)), this);
+ }
+
+ /**
+ * Schedule a new owner selection job. Cancelling any outstanding job if it has not been cancelled.
+ *
+ * @param entityPath
+ * @param allCandidates
+ */
+ public void scheduleOwnerSelection(YangInstanceIdentifier entityPath, Collection<String> allCandidates,
+ EntityOwnerSelectionStrategy strategy){
+ cancelOwnerSelectionTask(entityPath);
+
+ LOG.debug("{}: Scheduling owner selection after {} ms", persistenceId(), strategy.getSelectionDelayInMillis());
+
+ final Cancellable lastScheduledTask = context().system().scheduler().scheduleOnce(
+ FiniteDuration.apply(strategy.getSelectionDelayInMillis(), TimeUnit.MILLISECONDS)
+ , self(), new SelectOwner(entityPath, allCandidates, strategy)
+ , context().system().dispatcher(), self());
+
+ entityToScheduledOwnershipTask.put(entityPath, lastScheduledTask);
+ }
+
+ private void cancelOwnerSelectionTask(YangInstanceIdentifier entityPath){
+ final Cancellable lastScheduledTask = entityToScheduledOwnershipTask.get(entityPath);
+ if(lastScheduledTask != null && !lastScheduledTask.isCancelled()){
+ lastScheduledTask.cancel();
+ }
+ }
+
+ private String newOwner(String currentOwner, Collection<String> candidates, Map<String, Long> statistics, EntityOwnerSelectionStrategy ownerSelectionStrategy) {
+ Collection<String> viableCandidates = getViableCandidates(candidates);
+ if(viableCandidates.size() == 0){
+ return "";
+ }
+ return ownerSelectionStrategy.newOwner(currentOwner, viableCandidates);
+ }
+
+ private Collection<String> getViableCandidates(Collection<String> candidates) {
+ Collection<String> viableCandidates = new ArrayList<>();
+
+ for (String candidate : candidates) {
+ if (!downPeerMemberNames.contains(candidate)) {
+ viableCandidates.add(candidate);
+ }
+ }
+ return viableCandidates;
+ }
+
+ private String getCurrentOwner(YangInstanceIdentifier entityId) {
+ Optional<NormalizedNode<?, ?>> optionalEntityOwner = getDataStore().readNode(entityId.node(ENTITY_OWNER_QNAME));
+ if(optionalEntityOwner.isPresent()){
+ return optionalEntityOwner.get().getValue().toString();
+ }
+ return null;
+ }
+
+ private static interface EntityWalker {
+ void onEntity(MapEntryNode entityTypeNode, MapEntryNode entityNode);
+ }
+
+ public static Builder newBuilder() {
+ return new Builder();
+ }
+
+ static class Builder extends Shard.AbstractBuilder<Builder, EntityOwnershipShard> {
+ private MemberName localMemberName;
+ private EntityOwnerSelectionStrategyConfig ownerSelectionStrategyConfig;
+
+ protected Builder() {
+ super(EntityOwnershipShard.class);
+ }
+
+ Builder localMemberName(MemberName localMemberName) {
+ checkSealed();
+ this.localMemberName = localMemberName;
+ return this;
+ }