2 * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved.
3 * Copyright (c) 2015 Brocade Communications Systems, Inc. and others. All rights reserved.
5 * This program and the accompanying materials are made available under the
6 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
7 * and is available at http://www.eclipse.org/legal/epl-v10.html
10 package org.opendaylight.controller.cluster.raft;
12 import akka.actor.ActorRef;
13 import akka.actor.ActorSelection;
14 import akka.actor.PoisonPill;
15 import akka.japi.Procedure;
16 import com.google.common.annotations.VisibleForTesting;
17 import com.google.common.base.Optional;
18 import com.google.common.base.Preconditions;
19 import com.google.common.base.Verify;
20 import com.google.common.collect.Lists;
21 import java.io.Serializable;
22 import java.util.Collection;
23 import java.util.HashMap;
24 import java.util.List;
26 import java.util.Objects;
27 import java.util.concurrent.TimeUnit;
28 import javax.annotation.Nonnull;
29 import javax.annotation.Nullable;
30 import org.apache.commons.lang3.time.DurationFormatUtils;
31 import org.opendaylight.controller.cluster.DataPersistenceProvider;
32 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
33 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
34 import org.opendaylight.controller.cluster.PersistentDataProvider;
35 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
36 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
37 import org.opendaylight.controller.cluster.notifications.RoleChanged;
38 import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
39 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
40 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
41 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
42 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
43 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
44 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
45 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
46 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
47 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
48 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
49 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
50 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
51 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
52 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
53 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
54 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
55 import org.opendaylight.yangtools.concepts.Immutable;
56 import org.slf4j.Logger;
57 import org.slf4j.LoggerFactory;
60 * RaftActor encapsulates a state machine that needs to be kept synchronized
61 * in a cluster. It implements the RAFT algorithm as described in the paper
62 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
63 * In Search of an Understandable Consensus Algorithm</a>
65 * RaftActor has 3 states and each state has a certain behavior associated
66 * with it. A Raft actor can behave as,
69 * <li> A Follower (or) </li>
70 * <li> A Candidate </li>
74 * A RaftActor MUST be a Leader in order to accept requests from clients to
75 * change the state of it's encapsulated state machine. Once a RaftActor becomes
76 * a Leader it is also responsible for ensuring that all followers ultimately
77 * have the same log and therefore the same state machine as itself.
80 * The current behavior of a RaftActor determines how election for leadership
81 * is initiated and how peer RaftActors react to request for votes.
84 * Each RaftActor also needs to know the current election term. It uses this
85 * information for a couple of things. One is to simply figure out who it
86 * voted for in the last election. Another is to figure out if the message
87 * it received to update it's state is stale.
90 * The RaftActor uses akka-persistence to store it's replicated log.
91 * Furthermore through it's behaviors a Raft Actor determines
94 * <li> when a log entry should be persisted </li>
95 * <li> when a log entry should be applied to the state machine (and) </li>
96 * <li> when a snapshot should be saved </li>
99 public abstract class RaftActor extends AbstractUntypedPersistentActor {
101 private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
103 protected final Logger LOG = LoggerFactory.getLogger(getClass());
106 * This context should NOT be passed directly to any other actor it is
107 * only to be consumed by the RaftActorBehaviors
109 private final RaftActorContextImpl context;
111 private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
113 private final PersistentDataProvider persistentProvider;
115 private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
117 private RaftActorRecoverySupport raftRecovery;
119 private RaftActorSnapshotMessageSupport snapshotSupport;
121 private RaftActorServerConfigurationSupport serverConfigurationSupport;
123 private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
125 private boolean shuttingDown;
127 public RaftActor(String id, Map<String, String> peerAddresses,
128 Optional<ConfigParams> configParams, short payloadVersion) {
130 persistentProvider = new PersistentDataProvider(this);
131 delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
133 context = new RaftActorContextImpl(this.getSelf(),
134 this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
135 -1, -1, peerAddresses,
136 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
137 delegatingPersistenceProvider, LOG);
139 context.setPayloadVersion(payloadVersion);
140 context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
144 public void preStart() throws Exception {
145 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
146 context.getConfigParams().getJournalRecoveryLogBatchSize());
150 snapshotSupport = newRaftActorSnapshotMessageSupport();
151 serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
155 public void postStop() {
161 protected void handleRecover(Object message) {
162 if(raftRecovery == null) {
163 raftRecovery = newRaftActorRecoverySupport();
166 boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
167 if(recoveryComplete) {
168 onRecoveryComplete();
170 initializeBehavior();
174 if (context.getReplicatedLog().size() > 0) {
175 self().tell(new InitiateCaptureSnapshot(), self());
176 LOG.info("{}: Snapshot capture initiated after recovery", persistenceId());
178 LOG.info("{}: Snapshot capture NOT initiated after recovery, journal empty", persistenceId());
183 protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
184 return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
188 void initializeBehavior(){
189 changeCurrentBehavior(new Follower(context));
193 protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
194 final RaftActorBehavior currentBehavior = getCurrentBehavior();
195 if (currentBehavior != null) {
197 currentBehavior.close();
198 } catch (Exception e) {
199 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
203 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
204 setCurrentBehavior(newBehavior);
205 handleBehaviorChange(state, newBehavior);
209 protected void handleCommand(final Object message) {
210 if (serverConfigurationSupport.handleMessage(message, getSender())) {
213 if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
217 if (message instanceof ApplyState) {
218 ApplyState applyState = (ApplyState) message;
220 long elapsedTime = (System.nanoTime() - applyState.getStartTime());
221 if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
222 LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
223 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
226 if(LOG.isDebugEnabled()) {
227 LOG.debug("{}: Applying state for log index {} data {}",
228 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
229 applyState.getReplicatedLogEntry().getData());
232 applyState(applyState.getClientActor(), applyState.getIdentifier(),
233 applyState.getReplicatedLogEntry().getData());
235 if (!hasFollowers()) {
236 // for single node, the capture should happen after the apply state
237 // as we delete messages from the persistent journal which have made it to the snapshot
238 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
239 // and recovery shows data missing
240 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
242 context.getSnapshotManager().trimLog(context.getLastApplied());
245 } else if (message instanceof ApplyJournalEntries) {
246 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
247 if(LOG.isDebugEnabled()) {
248 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
251 persistence().persist(applyEntries, NoopProcedure.instance());
253 } else if (message instanceof FindLeader) {
255 new FindLeaderReply(getLeaderAddress()),
258 } else if(message instanceof GetOnDemandRaftState) {
259 onGetOnDemandRaftStats();
260 } else if(message instanceof InitiateCaptureSnapshot) {
262 } else if(message instanceof SwitchBehavior) {
263 switchBehavior(((SwitchBehavior) message));
264 } else if(message instanceof LeaderTransitioning) {
265 onLeaderTransitioning();
266 } else if(message instanceof Shutdown) {
268 } else if(message instanceof Runnable) {
269 ((Runnable)message).run();
271 // Processing the message may affect the state, hence we need to capture it
272 final RaftActorBehavior currentBehavior = getCurrentBehavior();
273 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
274 final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
275 switchBehavior(state, nextBehavior);
279 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
280 LOG.debug("{}: Initiating leader transfer", persistenceId());
282 if(leadershipTransferInProgress == null) {
283 leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
284 leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
286 public void onSuccess(ActorRef raftActorRef) {
287 leadershipTransferInProgress = null;
291 public void onFailure(ActorRef raftActorRef) {
292 leadershipTransferInProgress = null;
296 leadershipTransferInProgress.addOnComplete(onComplete);
297 leadershipTransferInProgress.init();
299 LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
300 leadershipTransferInProgress.addOnComplete(onComplete);
304 private void onShutDown() {
305 LOG.debug("{}: onShutDown", persistenceId());
313 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
314 if (currentBehavior.state() != RaftState.Leader) {
315 // For non-leaders shutdown is a no-op
316 self().tell(PoisonPill.getInstance(), self());
320 if (context.hasFollowers()) {
321 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
323 public void onSuccess(ActorRef raftActorRef) {
324 LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
325 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
329 public void onFailure(ActorRef raftActorRef) {
330 LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
331 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
335 pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
337 protected void doRun() {
338 self().tell(PoisonPill.getInstance(), self());
342 protected void doCancel() {
343 self().tell(PoisonPill.getInstance(), self());
349 private void onLeaderTransitioning() {
350 LOG.debug("{}: onLeaderTransitioning", persistenceId());
351 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
352 if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
353 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
354 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
358 private void switchBehavior(SwitchBehavior message) {
359 if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
360 RaftState newState = message.getNewState();
361 if( newState == RaftState.Leader || newState == RaftState.Follower) {
362 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
363 AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
364 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
366 LOG.warn("Switching to behavior : {} - not supported", newState);
371 private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
372 setCurrentBehavior(nextBehavior);
373 handleBehaviorChange(oldBehaviorState, nextBehavior);
377 RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
378 return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
381 private void onGetOnDemandRaftStats() {
382 // Debugging message to retrieve raft stats.
384 Map<String, String> peerAddresses = new HashMap<>();
385 for(String peerId: context.getPeerIds()) {
386 peerAddresses.put(peerId, context.getPeerAddress(peerId));
389 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
390 OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
391 .commitIndex(context.getCommitIndex())
392 .currentTerm(context.getTermInformation().getCurrentTerm())
393 .inMemoryJournalDataSize(replicatedLog().dataSize())
394 .inMemoryJournalLogSize(replicatedLog().size())
395 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
396 .lastApplied(context.getLastApplied())
397 .lastIndex(replicatedLog().lastIndex())
398 .lastTerm(replicatedLog().lastTerm())
399 .leader(getLeaderId())
400 .raftState(currentBehavior.state().toString())
401 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
402 .snapshotIndex(replicatedLog().getSnapshotIndex())
403 .snapshotTerm(replicatedLog().getSnapshotTerm())
404 .votedFor(context.getTermInformation().getVotedFor())
405 .peerAddresses(peerAddresses)
406 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
408 ReplicatedLogEntry lastLogEntry = replicatedLog().last();
409 if (lastLogEntry != null) {
410 builder.lastLogIndex(lastLogEntry.getIndex());
411 builder.lastLogTerm(lastLogEntry.getTerm());
414 if(getCurrentBehavior() instanceof AbstractLeader) {
415 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
416 Collection<String> followerIds = leader.getFollowerIds();
417 List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
418 for(String id: followerIds) {
419 final FollowerLogInformation info = leader.getFollower(id);
420 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
421 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity())));
424 builder.followerInfoList(followerInfoList);
427 sender().tell(builder.build(), self());
431 private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
432 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
434 if (oldBehavior != currentBehavior){
438 String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
439 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
441 // it can happen that the state has not changed but the leader has changed.
442 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
443 if(!Objects.equals(lastValidLeaderId, currentBehavior.getLeaderId()) ||
444 oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
445 if(roleChangeNotifier.isPresent()) {
446 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
447 currentBehavior.getLeaderPayloadVersion()), getSelf());
450 onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
452 if(leadershipTransferInProgress != null) {
453 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
456 serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
459 if (roleChangeNotifier.isPresent() &&
460 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
461 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
462 currentBehavior.state().name()), getSelf());
466 protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
467 return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
471 public long snapshotSequenceNr() {
472 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
473 // so that we can delete the persistent journal based on the saved sequence-number
474 // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
475 // was saved and not the number we saved.
476 // We would want to override it , by asking akka to use the last-sequence number known to us.
477 return context.getSnapshotManager().getLastSequenceNumber();
481 * When a derived RaftActor needs to persist something it must call
488 protected void persistData(final ActorRef clientActor, final String identifier,
489 final Payload data) {
491 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
492 context.getReplicatedLog().lastIndex() + 1,
493 context.getTermInformation().getCurrentTerm(), data);
495 if(LOG.isDebugEnabled()) {
496 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
499 final RaftActorContext raftContext = getRaftActorContext();
501 replicatedLog().appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
503 public void apply(ReplicatedLogEntry replicatedLogEntry) {
504 if (!hasFollowers()){
505 // Increment the Commit Index and the Last Applied values
506 raftContext.setCommitIndex(replicatedLogEntry.getIndex());
507 raftContext.setLastApplied(replicatedLogEntry.getIndex());
509 // Apply the state immediately.
510 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry), self());
512 // Send a ApplyJournalEntries message so that we write the fact that we applied
513 // the state to durable storage
514 self().tell(new ApplyJournalEntries(replicatedLogEntry.getIndex()), self());
516 } else if (clientActor != null) {
517 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
519 // Send message for replication
520 getCurrentBehavior().handleMessage(getSelf(),
521 new Replicate(clientActor, identifier, replicatedLogEntry));
527 private ReplicatedLog replicatedLog() {
528 return context.getReplicatedLog();
531 protected String getId() {
532 return context.getId();
536 void setCurrentBehavior(RaftActorBehavior behavior) {
537 context.setCurrentBehavior(behavior);
540 protected RaftActorBehavior getCurrentBehavior() {
541 return context.getCurrentBehavior();
545 * Derived actors can call the isLeader method to check if the current
546 * RaftActor is the Leader or not
548 * @return true it this RaftActor is a Leader false otherwise
550 protected boolean isLeader() {
551 return context.getId().equals(getCurrentBehavior().getLeaderId());
554 protected final boolean isLeaderActive() {
555 return getRaftState() != RaftState.IsolatedLeader && !shuttingDown &&
556 !isLeadershipTransferInProgress();
559 private boolean isLeadershipTransferInProgress() {
560 return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
564 * Derived actor can call getLeader if they need a reference to the Leader.
565 * This would be useful for example in forwarding a request to an actor
566 * which is the leader
568 * @return A reference to the leader if known, null otherwise
570 protected ActorSelection getLeader(){
571 String leaderAddress = getLeaderAddress();
573 if(leaderAddress == null){
577 return context.actorSelection(leaderAddress);
582 * @return the current leader's id
584 protected final String getLeaderId(){
585 return getCurrentBehavior().getLeaderId();
589 protected final RaftState getRaftState() {
590 return getCurrentBehavior().state();
593 protected Long getCurrentTerm(){
594 return context.getTermInformation().getCurrentTerm();
597 protected RaftActorContext getRaftActorContext() {
601 protected void updateConfigParams(ConfigParams configParams) {
603 // obtain the RaftPolicy for oldConfigParams and the updated one.
604 String oldRaftPolicy = context.getConfigParams().
605 getCustomRaftPolicyImplementationClass();
606 String newRaftPolicy = configParams.
607 getCustomRaftPolicyImplementationClass();
609 LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
610 oldRaftPolicy, newRaftPolicy);
611 context.setConfigParams(configParams);
612 if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
613 // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
614 // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
615 // avoids potential disruption. Otherwise, switch to Follower normally.
616 RaftActorBehavior behavior = getCurrentBehavior();
617 if (behavior != null && behavior.state() == RaftState.Follower) {
618 String previousLeaderId = behavior.getLeaderId();
619 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
621 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
623 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
625 initializeBehavior();
630 public final DataPersistenceProvider persistence() {
631 return delegatingPersistenceProvider.getDelegate();
634 public void setPersistence(DataPersistenceProvider provider) {
635 delegatingPersistenceProvider.setDelegate(provider);
638 protected void setPersistence(boolean persistent) {
640 setPersistence(new PersistentDataProvider(this));
642 setPersistence(new NonPersistentDataProvider() {
644 * The way snapshotting works is,
646 * <li> RaftActor calls createSnapshot on the Shard
647 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
648 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
649 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
650 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
651 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
655 public void saveSnapshot(Object o) {
656 // Make saving Snapshot successful
657 // Committing the snapshot here would end up calling commit in the creating state which would
658 // be a state violation. That's why now we send a message to commit the snapshot.
659 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
666 * setPeerAddress sets the address of a known peer at a later time.
668 * This is to account for situations where a we know that a peer
669 * exists but we do not know an address up-front. This may also be used in
670 * situations where a known peer starts off in a different location and we
671 * need to change it's address
673 * Note that if the peerId does not match the list of peers passed to
674 * this actor during construction an IllegalStateException will be thrown.
679 protected void setPeerAddress(String peerId, String peerAddress){
680 context.setPeerAddress(peerId, peerAddress);
684 * The applyState method will be called by the RaftActor when some data
685 * needs to be applied to the actor's state
687 * @param clientActor A reference to the client who sent this message. This
688 * is the same reference that was passed to persistData
689 * by the derived actor. clientActor may be null when
690 * the RaftActor is behaving as a follower or during
692 * @param identifier The identifier of the persisted data. This is also
693 * the same identifier that was passed to persistData by
694 * the derived actor. identifier may be null when
695 * the RaftActor is behaving as a follower or during
697 * @param data A piece of data that was persisted by the persistData call.
698 * This should NEVER be null.
700 protected abstract void applyState(ActorRef clientActor, String identifier,
704 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
707 protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
710 * This method is called when recovery is complete.
712 protected abstract void onRecoveryComplete();
715 * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
718 protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
721 * This method will be called by the RaftActor when the state of the
722 * RaftActor changes. The derived actor can then use methods like
723 * isLeader or getLeader to do something useful
725 protected abstract void onStateChanged();
728 * Notifier Actor for this RaftActor to notify when a role change happens
729 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
731 protected abstract Optional<ActorRef> getRoleChangeNotifier();
734 * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
735 * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
736 * work prior to performing the operation. On completion of any work, the run method must be called on the
737 * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
738 * this actor's thread dispatcher as as it modifies internal state.
740 * The default implementation immediately runs the operation.
742 * @param operation the operation to run
744 protected void pauseLeader(Runnable operation) {
748 protected void onLeaderChanged(String oldLeader, String newLeader) {
752 private String getLeaderAddress(){
754 return getSelf().path().toString();
756 String leaderId = getLeaderId();
757 if (leaderId == null) {
760 String peerAddress = context.getPeerAddress(leaderId);
761 if(LOG.isDebugEnabled()) {
762 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
763 persistenceId(), leaderId, peerAddress);
769 protected boolean hasFollowers(){
770 return getRaftActorContext().hasFollowers();
773 private void captureSnapshot() {
774 SnapshotManager snapshotManager = context.getSnapshotManager();
776 if (!snapshotManager.isCapturing()) {
777 final long idx = getCurrentBehavior().getReplicatedToAllIndex();
778 LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
779 replicatedLog().last(), idx);
781 snapshotManager.capture(replicatedLog().last(), idx);
786 * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
787 * in which case we need to step down.
789 void becomeNonVoting() {
791 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
793 public void onSuccess(ActorRef raftActorRef) {
794 LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
795 ensureFollowerState();
799 public void onFailure(ActorRef raftActorRef) {
800 LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
801 ensureFollowerState();
804 private void ensureFollowerState() {
805 // Whether or not leadership transfer succeeded, we have to step down as leader and
806 // switch to Follower so ensure that.
807 if (getRaftState() != RaftState.Follower) {
808 initializeBehavior();
816 * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
817 * whose type for fromIndex is long instead of int. This class was kept for backwards
818 * compatibility with Helium.
820 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
821 @SuppressWarnings("serial")
823 static class DeleteEntries implements Serializable {
824 private final int fromIndex;
826 public DeleteEntries(int fromIndex) {
827 this.fromIndex = fromIndex;
830 public int getFromIndex() {
836 * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
837 * which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
839 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
840 @SuppressWarnings("serial")
842 static class UpdateElectionTerm implements Serializable {
843 private final long currentTerm;
844 private final String votedFor;
846 public UpdateElectionTerm(long currentTerm, String votedFor) {
847 this.currentTerm = currentTerm;
848 this.votedFor = votedFor;
851 public long getCurrentTerm() {
855 public String getVotedFor() {
861 * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
863 private static abstract class BehaviorState implements Immutable {
864 @Nullable abstract RaftActorBehavior getBehavior();
865 @Nullable abstract String getLastValidLeaderId();
866 @Nullable abstract short getLeaderPayloadVersion();
870 * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
872 private static final class SimpleBehaviorState extends BehaviorState {
873 private final RaftActorBehavior behavior;
874 private final String lastValidLeaderId;
875 private final short leaderPayloadVersion;
877 SimpleBehaviorState(final String lastValidLeaderId, final RaftActorBehavior behavior) {
878 this.lastValidLeaderId = lastValidLeaderId;
879 this.behavior = Preconditions.checkNotNull(behavior);
880 this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
884 RaftActorBehavior getBehavior() {
889 String getLastValidLeaderId() {
890 return lastValidLeaderId;
894 short getLeaderPayloadVersion() {
895 return leaderPayloadVersion;
900 * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
901 * An instance is created for each RaftActor. It has two functions:
902 * - it keeps track of the last leader ID we have encountered since we have been created
903 * - it creates state capture needed to transition from one behavior to the next
905 private static final class BehaviorStateTracker {
907 * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
908 * allowed before we receive the first message, we know the leader ID to be null.
910 private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
912 RaftActorBehavior getBehavior() {
917 String getLastValidLeaderId() {
922 short getLeaderPayloadVersion() {
927 private String lastValidLeaderId;
929 BehaviorState capture(final RaftActorBehavior behavior) {
930 if (behavior == null) {
931 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
932 return NULL_BEHAVIOR_STATE;
935 final String leaderId = behavior.getLeaderId();
936 if (leaderId != null) {
937 lastValidLeaderId = leaderId;
940 return new SimpleBehaviorState(lastValidLeaderId, behavior);