+ @Override
+ protected void onRecoveryComplete() {
+ super.onRecoveryComplete();
+
+ new CandidateListChangeListener(getSelf(), persistenceId()).init(getDataStore());
+ new EntityOwnerChangeListener(localMemberName, listenerSupport).init(getDataStore());
+ }
+
+ @Override
+ public void handleNonRaftCommand(final Object message) {
+ if (message instanceof RegisterCandidateLocal) {
+ onRegisterCandidateLocal((RegisterCandidateLocal) message);
+ } else if (message instanceof UnregisterCandidateLocal) {
+ onUnregisterCandidateLocal((UnregisterCandidateLocal) message);
+ } else if (message instanceof CandidateAdded) {
+ onCandidateAdded((CandidateAdded) message);
+ } else if (message instanceof CandidateRemoved) {
+ onCandidateRemoved((CandidateRemoved) message);
+ } else if (message instanceof PeerDown) {
+ onPeerDown((PeerDown) message);
+ } else if (message instanceof PeerUp) {
+ onPeerUp((PeerUp) message);
+ } else if (message instanceof RegisterListenerLocal) {
+ onRegisterListenerLocal((RegisterListenerLocal) message);
+ } else if (message instanceof UnregisterListenerLocal) {
+ onUnregisterListenerLocal((UnregisterListenerLocal) message);
+ } else if (message instanceof SelectOwner) {
+ onSelectOwner((SelectOwner) message);
+ } else if (message instanceof RemoveAllCandidates) {
+ onRemoveAllCandidates((RemoveAllCandidates) message);
+ } else if (!commitCoordinator.handleMessage(message, this)) {
+ super.handleNonRaftCommand(message);
+ }
+ }
+
+ private void onRemoveAllCandidates(final RemoveAllCandidates message) {
+ LOG.debug("{}: onRemoveAllCandidates: {}", persistenceId(), message);
+
+ removeCandidateFromEntities(message.getMemberName());
+ }
+
+ private void onSelectOwner(final SelectOwner selectOwner) {
+ LOG.debug("{}: onSelectOwner: {}", persistenceId(), selectOwner);
+
+ String currentOwner = getCurrentOwner(selectOwner.getEntityPath());
+ if (Strings.isNullOrEmpty(currentOwner)) {
+ writeNewOwner(selectOwner.getEntityPath(), newOwner(currentOwner, selectOwner.getAllCandidates(),
+ selectOwner.getOwnerSelectionStrategy()));
+
+ Cancellable cancellable = entityToScheduledOwnershipTask.get(selectOwner.getEntityPath());
+ if (cancellable != null) {
+ if (!cancellable.isCancelled()) {
+ cancellable.cancel();
+ }
+ entityToScheduledOwnershipTask.remove(selectOwner.getEntityPath());
+ }
+ }
+ }
+
+ private void onRegisterCandidateLocal(final RegisterCandidateLocal registerCandidate) {
+ LOG.debug("{}: onRegisterCandidateLocal: {}", persistenceId(), registerCandidate);
+
+ NormalizedNode<?, ?> entityOwners = entityOwnersWithCandidate(registerCandidate.getEntity().getType(),
+ registerCandidate.getEntity().getIdentifier(), localMemberName.getName());
+ commitCoordinator.commitModification(new MergeModification(ENTITY_OWNERS_PATH, entityOwners), this);
+
+ getSender().tell(SuccessReply.INSTANCE, getSelf());
+ }
+
+ private void onUnregisterCandidateLocal(final UnregisterCandidateLocal unregisterCandidate) {
+ LOG.debug("{}: onUnregisterCandidateLocal: {}", persistenceId(), unregisterCandidate);
+
+ DOMEntity entity = unregisterCandidate.getEntity();
+ YangInstanceIdentifier candidatePath = candidatePath(entity.getType(), entity.getIdentifier(),
+ localMemberName.getName());
+ commitCoordinator.commitModification(new DeleteModification(candidatePath), this);
+
+ getSender().tell(SuccessReply.INSTANCE, getSelf());
+ }
+
+ 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) -> {
+ java.util.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;
+
+ java.util.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(final 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) {
+ possiblyRemoveAllInitialCandidates(leader);
+
+ 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());
+ }
+ }
+ }
+
+ void possiblyRemoveAllInitialCandidates(final ActorSelection leader) {
+ // The following handles removing all candidates on startup when re-joining with a remote leader. When a
+ // follower is detected as down, the leader will re-assign new owners to entities that were owned by the
+ // down member but doesn't remove the down member as a candidate, as the down node may actually be isolated
+ // and still running. Therefore on startup we send an initial message to the remote leader to remove any
+ // potential stale candidates we had previously registered, as it's possible a candidate may not be
+ // registered by a client in the new incarnation. We have to send the RemoveAllCandidates message prior to any
+ // pending registrations.
+ if (removeAllInitialCandidates && leader != null) {
+ removeAllInitialCandidates = false;
+ if (!isLeader()) {
+ LOG.debug("{} - got new leader {} on startup - sending RemoveAllCandidates", persistenceId(), leader);
+
+ leader.tell(new RemoveAllCandidates(localMemberName), ActorRef.noSender());
+ }
+ }
+ }
+
+ boolean hasLeader() {
+ return getLeader() != null && (!isLeader() || isLeaderActive());
+ }
+
+ /**
+ * 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:
+ case PreLeader:
+ return false;
+ case IsolatedLeader:
+ return true;
+ default:
+ throw new IllegalStateException("Unsupported RAFT state " + state);
+ }
+ }
+
+ private void notifyAllListeners() {
+ searchForEntities((entityTypeNode, entityNode) -> {
+ java.util.Optional<DataContainerChild<?, ?>> possibleType = entityTypeNode.getChild(ENTITY_TYPE_NODE_ID);
+ if (possibleType.isPresent()) {
+ final boolean hasOwner;
+ final boolean isOwner;
+
+ java.util.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() {
+ 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);
+
+ super.onStateChanged();
+ }
+
+ @Override
+ protected void onLeaderChanged(final String oldLeader, final String newLeader) {
+ boolean isLeader = isLeader();
+ LOG.debug("{}: onLeaderChanged: oldLeader: {}, newLeader: {}, isLeader: {}", persistenceId(), oldLeader,
+ newLeader, isLeader);
+
+ if (isLeader) {
+
+ // Re-initialize the downPeerMemberNames from the current akka Cluster state. The previous leader, if any,
+ // is most likely down however it's possible we haven't received the PeerDown message yet.
+ initializeDownPeerMemberNamesFromClusterState();
+
+ // 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();
+
+ // Re-assign owners for all members that are known to be down. In a cluster which has greater than
+ // 3 nodes it is possible for some node beside the leader being down when the leadership transitions
+ // it makes sense to use this event to re-assign owners for those downed nodes.
+ Set<String> ownedBy = new HashSet<>(downPeerMemberNames.size() + 1);
+ for (MemberName downPeerName : downPeerMemberNames) {
+ ownedBy.add(downPeerName.getName());
+ }
+
+ // Also try to assign owners for entities that have no current owner. See explanation in onPeerUp.
+ ownedBy.add("");
+ selectNewOwnerForEntitiesOwnedBy(ownedBy);
+ } 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);
+ }
+
+ super.onLeaderChanged(oldLeader, newLeader);
+ }
+
+ @Override
+ protected void onVotingStateChangeComplete() {
+ // Re-evaluate ownership for all entities - if a member changed from voting to non-voting it should lose
+ // ownership and vice versa it now is a candidate to become owner.
+ final List<Modification> modifications = new ArrayList<>();
+ searchForEntities((entityTypeNode, entityNode) -> {
+ YangInstanceIdentifier entityPath = YangInstanceIdentifier.builder(ENTITY_TYPES_PATH)
+ .node(entityTypeNode.getIdentifier()).node(ENTITY_NODE_ID).node(entityNode.getIdentifier())
+ .node(ENTITY_OWNER_NODE_ID).build();
+
+ java.util.Optional<String> possibleOwner =
+ entityNode.getChild(ENTITY_OWNER_NODE_ID).map(node -> node.getValue().toString());
+ String newOwner = newOwner(possibleOwner.orElse(null), getCandidateNames(entityNode),
+ getEntityOwnerElectionStrategy(entityPath));
+
+ if (!newOwner.equals(possibleOwner.orElse(""))) {
+ modifications.add(new WriteModification(entityPath,
+ ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)));
+ }
+ });
+
+ commitCoordinator.commitModifications(modifications, this);
+ }
+
+ private void initializeDownPeerMemberNamesFromClusterState() {
+ java.util.Optional<Cluster> cluster = getRaftActorContext().getCluster();
+ if (!cluster.isPresent()) {
+ return;
+ }
+
+ CurrentClusterState state = cluster.get().state();
+ Set<Member> unreachable = state.getUnreachable();
+
+ LOG.debug(
+ "{}: initializeDownPeerMemberNamesFromClusterState - current downPeerMemberNames: {}, unreachable: {}",
+ persistenceId(), downPeerMemberNames, unreachable);
+
+ downPeerMemberNames.clear();
+ for (Member m: unreachable) {
+ downPeerMemberNames.add(MemberName.forName(m.getRoles().iterator().next()));
+ }
+
+ for (Member m: state.getMembers()) {
+ if (m.status() != MemberStatus.up() && m.status() != MemberStatus.weaklyUp()) {
+ LOG.debug("{}: Adding down member with status {}", persistenceId(), m.status());
+ downPeerMemberNames.add(MemberName.forName(m.getRoles().iterator().next()));
+ }
+ }
+
+ LOG.debug("{}: new downPeerMemberNames: {}", persistenceId(), downPeerMemberNames);
+ }
+
+ private void onCandidateRemoved(final CandidateRemoved message) {
+ LOG.debug("{}: onCandidateRemoved: {}", persistenceId(), message);
+
+ if (isLeader()) {
+ String currentOwner = getCurrentOwner(message.getEntityPath());
+ writeNewOwner(message.getEntityPath(),
+ newOwner(currentOwner, message.getRemainingCandidates(),
+ getEntityOwnerElectionStrategy(message.getEntityPath())));
+ }
+ }
+
+ private EntityOwnerSelectionStrategy getEntityOwnerElectionStrategy(final YangInstanceIdentifier entityPath) {
+ final String entityType = EntityOwnersModel.entityTypeFromEntityPath(entityPath);
+ return strategyConfig.createStrategy(entityType, entityOwnershipStatistics.byEntityType(entityType));
+ }
+
+ private void onCandidateAdded(final 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(MemberName.forName(message.getNewCandidate()));
+
+ final String currentOwner = getCurrentOwner(message.getEntityPath());
+ final EntityOwnerSelectionStrategy strategy = getEntityOwnerElectionStrategy(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 = getRaftActorContext().getPeerIds().size() - downPeerMemberNames.size() + 1;
+
+ LOG.debug("{}: Using strategy {} to select owner, currentOwner = {}", persistenceId(), strategy, currentOwner);
+
+ if (strategy.getSelectionDelayInMillis() == 0L) {
+ writeNewOwner(message.getEntityPath(), newOwner(currentOwner, message.getAllCandidates(),
+ 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(),
+ strategy));
+ } else {
+ scheduleOwnerSelection(message.getEntityPath(), message.getAllCandidates(), strategy);
+ }
+ }
+
+ private void onPeerDown(final PeerDown peerDown) {
+ LOG.info("{}: onPeerDown: {}", persistenceId(), peerDown);
+
+ MemberName downMemberName = peerDown.getMemberName();
+ if (downPeerMemberNames.add(downMemberName) && isLeader()) {
+ // Select new owners for entities owned by the down peer and which have other candidates. For an entity for
+ // which the down peer is the only candidate, we leave it as the owner and don't clear it. This is done to
+ // handle the case where the peer member process is actually still running but the node is partitioned.
+ // When the partition is healed, the peer just remains as the owner. If the peer process actually restarted,
+ // it will first remove all its candidates on startup. If another candidate is registered during the time
+ // the peer is down, the new candidate will be selected as the new owner.
+
+ selectNewOwnerForEntitiesOwnedBy(ImmutableSet.of(downMemberName.getName()));
+ }
+ }
+
+ private void selectNewOwnerForEntitiesOwnedBy(final Set<String> ownedBy) {
+ final List<Modification> modifications = new ArrayList<>();
+ searchForEntitiesOwnedBy(ownedBy, (entityTypeNode, entityNode) -> {
+ YangInstanceIdentifier entityPath = YangInstanceIdentifier.builder(ENTITY_TYPES_PATH)
+ .node(entityTypeNode.getIdentifier()).node(ENTITY_NODE_ID).node(entityNode.getIdentifier())
+ .node(ENTITY_OWNER_NODE_ID).build();
+ String newOwner = newOwner(getCurrentOwner(entityPath), getCandidateNames(entityNode),
+ getEntityOwnerElectionStrategy(entityPath));
+
+ if (!newOwner.isEmpty()) {
+ LOG.debug("{}: Found entity {}, writing new owner {}", persistenceId(), entityPath, newOwner);
+
+ modifications.add(new WriteModification(entityPath,
+ ImmutableNodes.leafNode(ENTITY_OWNER_NODE_ID, newOwner)));
+
+ } else {
+ LOG.debug("{}: Found entity {} but no other candidates - not clearing owner", persistenceId(),
+ entityPath);
+ }
+ });
+
+ commitCoordinator.commitModifications(modifications, this);
+ }
+
+ private void onPeerUp(final PeerUp peerUp) {
+ LOG.debug("{}: onPeerUp: {}", persistenceId(), peerUp);
+
+ 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 known yet when a prior state or
+ // leader change occurred.
+ commitCoordinator.onStateChanged(this, isLeader());
+
+ if (isLeader()) {
+ // Try to assign owners for entities that have no current owner. It's possible the peer that is now up
+ // had previously registered as a candidate and was the only candidate but the owner write tx couldn't be
+ // committed due to a leader change. Eg, the leader is able to successfully commit the candidate add tx but
+ // becomes isolated before it can commit the owner change and switches to follower. The majority partition
+ // with a new leader has the candidate but the entity has no owner. When the partition is healed and the
+ // previously isolated leader reconnects, we'll receive onPeerUp and, if there's still no owner, the
+ // previous leader will gain ownership.
+ selectNewOwnerForEntitiesOwnedBy(ImmutableSet.of(""));
+ }
+ }
+
+ private static Collection<String> getCandidateNames(final MapEntryNode entity) {
+ Collection<MapEntryNode> candidates = ((MapNode)entity.getChild(CANDIDATE_NODE_ID).get()).getValue();
+ Collection<String> candidateNames = new ArrayList<>(candidates.size());
+ for (MapEntryNode candidate: candidates) {
+ candidateNames.add(candidate.getChild(CANDIDATE_NAME_NODE_ID).get().getValue().toString());
+ }
+
+ return candidateNames;
+ }
+
+ private void searchForEntitiesOwnedBy(final Set<String> ownedBy, final EntityWalker walker) {
+ LOG.debug("{}: Searching for entities owned by {}", persistenceId(), ownedBy);
+
+ searchForEntities((entityTypeNode, entityNode) -> {
+ java.util.Optional<DataContainerChild<? extends PathArgument, ?>> possibleOwner =
+ entityNode.getChild(ENTITY_OWNER_NODE_ID);
+ String currentOwner = possibleOwner.isPresent() ? possibleOwner.get().getValue().toString() : "";
+ if (ownedBy.contains(currentOwner)) {
+ walker.onEntity(entityTypeNode, entityNode);
+ }
+ });