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 com.google.common.annotations.VisibleForTesting;
16 import com.google.common.base.Optional;
17 import com.google.common.base.Preconditions;
18 import com.google.common.base.Verify;
19 import com.google.common.collect.Lists;
20 import java.io.Serializable;
21 import java.util.Collection;
22 import java.util.HashMap;
23 import java.util.List;
25 import java.util.Objects;
26 import java.util.concurrent.TimeUnit;
27 import javax.annotation.Nonnull;
28 import javax.annotation.Nullable;
29 import org.apache.commons.lang3.time.DurationFormatUtils;
30 import org.opendaylight.controller.cluster.DataPersistenceProvider;
31 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
32 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
33 import org.opendaylight.controller.cluster.PersistentDataProvider;
34 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
35 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
36 import org.opendaylight.controller.cluster.notifications.RoleChanged;
37 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
38 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
39 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
40 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
41 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
44 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
45 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
48 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
49 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
50 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
51 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
52 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
53 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
54 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
55 import org.opendaylight.yangtools.concepts.Identifier;
56 import org.opendaylight.yangtools.concepts.Immutable;
57 import org.slf4j.Logger;
58 import org.slf4j.LoggerFactory;
61 * RaftActor encapsulates a state machine that needs to be kept synchronized
62 * in a cluster. It implements the RAFT algorithm as described in the paper
63 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
64 * In Search of an Understandable Consensus Algorithm</a>
66 * RaftActor has 3 states and each state has a certain behavior associated
67 * with it. A Raft actor can behave as,
70 * <li> A Follower (or) </li>
71 * <li> A Candidate </li>
75 * A RaftActor MUST be a Leader in order to accept requests from clients to
76 * change the state of it's encapsulated state machine. Once a RaftActor becomes
77 * a Leader it is also responsible for ensuring that all followers ultimately
78 * have the same log and therefore the same state machine as itself.
81 * The current behavior of a RaftActor determines how election for leadership
82 * is initiated and how peer RaftActors react to request for votes.
85 * Each RaftActor also needs to know the current election term. It uses this
86 * information for a couple of things. One is to simply figure out who it
87 * voted for in the last election. Another is to figure out if the message
88 * it received to update it's state is stale.
91 * The RaftActor uses akka-persistence to store it's replicated log.
92 * Furthermore through it's behaviors a Raft Actor determines
95 * <li> when a log entry should be persisted </li>
96 * <li> when a log entry should be applied to the state machine (and) </li>
97 * <li> when a snapshot should be saved </li>
100 public abstract class RaftActor extends AbstractUntypedPersistentActor {
102 private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
104 protected final Logger LOG = LoggerFactory.getLogger(getClass());
107 * This context should NOT be passed directly to any other actor it is
108 * only to be consumed by the RaftActorBehaviors
110 private final RaftActorContextImpl context;
112 private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
114 private final PersistentDataProvider persistentProvider;
116 private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
118 private RaftActorRecoverySupport raftRecovery;
120 private RaftActorSnapshotMessageSupport snapshotSupport;
122 private RaftActorServerConfigurationSupport serverConfigurationSupport;
124 private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
126 private boolean shuttingDown;
128 public RaftActor(String id, Map<String, String> peerAddresses,
129 Optional<ConfigParams> configParams, short payloadVersion) {
131 persistentProvider = new PersistentDataProvider(this);
132 delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
134 context = new RaftActorContextImpl(this.getSelf(),
135 this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
136 -1, -1, peerAddresses,
137 configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl(),
138 delegatingPersistenceProvider, LOG);
140 context.setPayloadVersion(payloadVersion);
141 context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
145 public void preStart() throws Exception {
146 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
147 context.getConfigParams().getJournalRecoveryLogBatchSize());
151 snapshotSupport = newRaftActorSnapshotMessageSupport();
152 serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
156 public void postStop() {
162 protected void handleRecover(Object message) {
163 if(raftRecovery == null) {
164 raftRecovery = newRaftActorRecoverySupport();
167 boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
168 if(recoveryComplete) {
169 onRecoveryComplete();
171 initializeBehavior();
177 protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
178 return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
182 void initializeBehavior(){
183 changeCurrentBehavior(new Follower(context));
187 protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
188 final RaftActorBehavior currentBehavior = getCurrentBehavior();
189 if (currentBehavior != null) {
191 currentBehavior.close();
192 } catch (Exception e) {
193 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
197 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
198 setCurrentBehavior(newBehavior);
199 handleBehaviorChange(state, newBehavior);
203 * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
204 * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
205 * implementation for messages which they do not handle
207 * @param message Incoming command message
209 protected void handleNonRaftCommand(final Object message) {
214 * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
215 * {@link #handleNonRaftCommand(Object)} instead.
219 // FIXME: make this method final once our unit tests do not need to override it
220 protected void handleCommand(final Object message) {
221 if (serverConfigurationSupport.handleMessage(message, getSender())) {
224 if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
228 if (message instanceof ApplyState) {
229 ApplyState applyState = (ApplyState) message;
231 long startTime = System.nanoTime();
233 if(LOG.isDebugEnabled()) {
234 LOG.debug("{}: Applying state for log index {} data {}",
235 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
236 applyState.getReplicatedLogEntry().getData());
239 if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
240 applyState(applyState.getClientActor(), applyState.getIdentifier(),
241 applyState.getReplicatedLogEntry().getData());
244 long elapsedTime = System.nanoTime() - startTime;
245 if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
246 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
247 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
250 if (!hasFollowers()) {
251 // for single node, the capture should happen after the apply state
252 // as we delete messages from the persistent journal which have made it to the snapshot
253 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
254 // and recovery shows data missing
255 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
257 context.getSnapshotManager().trimLog(context.getLastApplied());
260 } else if (message instanceof ApplyJournalEntries) {
261 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
262 if(LOG.isDebugEnabled()) {
263 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
266 persistence().persist(applyEntries, NoopProcedure.instance());
268 } else if (message instanceof FindLeader) {
270 new FindLeaderReply(getLeaderAddress()),
273 } else if(message instanceof GetOnDemandRaftState) {
274 onGetOnDemandRaftStats();
275 } else if(message instanceof InitiateCaptureSnapshot) {
277 } else if(message instanceof SwitchBehavior) {
278 switchBehavior((SwitchBehavior) message);
279 } else if(message instanceof LeaderTransitioning) {
280 onLeaderTransitioning();
281 } else if(message instanceof Shutdown) {
283 } else if(message instanceof Runnable) {
284 ((Runnable)message).run();
285 } else if(message instanceof NoopPayload) {
286 persistData(null, null, (NoopPayload)message);
288 // Processing the message may affect the state, hence we need to capture it
289 final RaftActorBehavior currentBehavior = getCurrentBehavior();
290 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
292 // A behavior indicates that it processed the change by returning a reference to the next behavior
293 // to be used. A null return indicates it has not processed the message and we should be passing it to
294 // the subclass for handling.
295 final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
296 if (nextBehavior != null) {
297 switchBehavior(state, nextBehavior);
299 handleNonRaftCommand(message);
304 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
305 LOG.debug("{}: Initiating leader transfer", persistenceId());
307 if(leadershipTransferInProgress == null) {
308 leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
309 leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
311 public void onSuccess(ActorRef raftActorRef) {
312 leadershipTransferInProgress = null;
316 public void onFailure(ActorRef raftActorRef) {
317 leadershipTransferInProgress = null;
321 leadershipTransferInProgress.addOnComplete(onComplete);
322 leadershipTransferInProgress.init();
324 LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
325 leadershipTransferInProgress.addOnComplete(onComplete);
329 private void onShutDown() {
330 LOG.debug("{}: onShutDown", persistenceId());
338 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
339 if (currentBehavior.state() != RaftState.Leader) {
340 // For non-leaders shutdown is a no-op
341 self().tell(PoisonPill.getInstance(), self());
345 if (context.hasFollowers()) {
346 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
348 public void onSuccess(ActorRef raftActorRef) {
349 LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
350 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
354 public void onFailure(ActorRef raftActorRef) {
355 LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
356 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
360 pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
362 protected void doRun() {
363 self().tell(PoisonPill.getInstance(), self());
367 protected void doCancel() {
368 self().tell(PoisonPill.getInstance(), self());
374 private void onLeaderTransitioning() {
375 LOG.debug("{}: onLeaderTransitioning", persistenceId());
376 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
377 if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
378 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
379 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
383 private void switchBehavior(SwitchBehavior message) {
384 if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
385 RaftState newState = message.getNewState();
386 if( newState == RaftState.Leader || newState == RaftState.Follower) {
387 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
388 AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
389 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
391 LOG.warn("Switching to behavior : {} - not supported", newState);
396 private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
397 setCurrentBehavior(nextBehavior);
398 handleBehaviorChange(oldBehaviorState, nextBehavior);
402 RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
403 return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
406 private void onGetOnDemandRaftStats() {
407 // Debugging message to retrieve raft stats.
409 Map<String, String> peerAddresses = new HashMap<>();
410 Map<String, Boolean> peerVotingStates = new HashMap<>();
411 for(PeerInfo info: context.getPeers()) {
412 peerVotingStates.put(info.getId(), info.isVoting());
413 peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
416 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
417 OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
418 .commitIndex(context.getCommitIndex())
419 .currentTerm(context.getTermInformation().getCurrentTerm())
420 .inMemoryJournalDataSize(replicatedLog().dataSize())
421 .inMemoryJournalLogSize(replicatedLog().size())
422 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
423 .lastApplied(context.getLastApplied())
424 .lastIndex(replicatedLog().lastIndex())
425 .lastTerm(replicatedLog().lastTerm())
426 .leader(getLeaderId())
427 .raftState(currentBehavior.state().toString())
428 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
429 .snapshotIndex(replicatedLog().getSnapshotIndex())
430 .snapshotTerm(replicatedLog().getSnapshotTerm())
431 .votedFor(context.getTermInformation().getVotedFor())
432 .isVoting(context.isVotingMember())
433 .peerAddresses(peerAddresses)
434 .peerVotingStates(peerVotingStates)
435 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
437 ReplicatedLogEntry lastLogEntry = replicatedLog().last();
438 if (lastLogEntry != null) {
439 builder.lastLogIndex(lastLogEntry.getIndex());
440 builder.lastLogTerm(lastLogEntry.getTerm());
443 if(getCurrentBehavior() instanceof AbstractLeader) {
444 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
445 Collection<String> followerIds = leader.getFollowerIds();
446 List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
447 for(String id: followerIds) {
448 final FollowerLogInformation info = leader.getFollower(id);
449 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
450 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
451 context.getPeerInfo(info.getId()).isVoting()));
454 builder.followerInfoList(followerInfoList);
457 sender().tell(builder.build(), self());
461 private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
462 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
464 if (oldBehavior != currentBehavior){
468 String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
469 String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
470 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
472 // it can happen that the state has not changed but the leader has changed.
473 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
474 if(!Objects.equals(lastLeaderId, currentBehavior.getLeaderId()) ||
475 oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
476 if(roleChangeNotifier.isPresent()) {
477 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
478 currentBehavior.getLeaderPayloadVersion()), getSelf());
481 onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
483 if(leadershipTransferInProgress != null) {
484 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
487 serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
490 if (roleChangeNotifier.isPresent() &&
491 (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
492 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
493 currentBehavior.state().name()), getSelf());
497 protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
498 return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
502 public long snapshotSequenceNr() {
503 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
504 // so that we can delete the persistent journal based on the saved sequence-number
505 // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
506 // was saved and not the number we saved.
507 // We would want to override it , by asking akka to use the last-sequence number known to us.
508 return context.getSnapshotManager().getLastSequenceNumber();
512 * When a derived RaftActor needs to persist something it must call
519 protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
521 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
522 context.getReplicatedLog().lastIndex() + 1,
523 context.getTermInformation().getCurrentTerm(), data);
525 if(LOG.isDebugEnabled()) {
526 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
529 final RaftActorContext raftContext = getRaftActorContext();
531 replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
532 if (!hasFollowers()){
533 // Increment the Commit Index and the Last Applied values
534 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
535 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
537 // Apply the state immediately.
538 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
540 // Send a ApplyJournalEntries message so that we write the fact that we applied
541 // the state to durable storage
542 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
545 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
547 // Send message for replication
548 getCurrentBehavior().handleMessage(getSelf(),
549 new Replicate(clientActor, identifier, replicatedLogEntry1));
554 private ReplicatedLog replicatedLog() {
555 return context.getReplicatedLog();
558 protected String getId() {
559 return context.getId();
563 void setCurrentBehavior(RaftActorBehavior behavior) {
564 context.setCurrentBehavior(behavior);
567 protected RaftActorBehavior getCurrentBehavior() {
568 return context.getCurrentBehavior();
572 * Derived actors can call the isLeader method to check if the current
573 * RaftActor is the Leader or not
575 * @return true it this RaftActor is a Leader false otherwise
577 protected boolean isLeader() {
578 return context.getId().equals(getCurrentBehavior().getLeaderId());
581 protected final boolean isLeaderActive() {
582 return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader &&
583 !shuttingDown && !isLeadershipTransferInProgress();
586 private boolean isLeadershipTransferInProgress() {
587 return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
591 * Derived actor can call getLeader if they need a reference to the Leader.
592 * This would be useful for example in forwarding a request to an actor
593 * which is the leader
595 * @return A reference to the leader if known, null otherwise
597 protected ActorSelection getLeader(){
598 String leaderAddress = getLeaderAddress();
600 if(leaderAddress == null){
604 return context.actorSelection(leaderAddress);
609 * @return the current leader's id
611 protected final String getLeaderId(){
612 return getCurrentBehavior().getLeaderId();
616 protected final RaftState getRaftState() {
617 return getCurrentBehavior().state();
620 protected Long getCurrentTerm(){
621 return context.getTermInformation().getCurrentTerm();
624 protected RaftActorContext getRaftActorContext() {
628 protected void updateConfigParams(ConfigParams configParams) {
630 // obtain the RaftPolicy for oldConfigParams and the updated one.
631 String oldRaftPolicy = context.getConfigParams().
632 getCustomRaftPolicyImplementationClass();
633 String newRaftPolicy = configParams.
634 getCustomRaftPolicyImplementationClass();
636 LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
637 oldRaftPolicy, newRaftPolicy);
638 context.setConfigParams(configParams);
639 if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
640 // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
641 // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
642 // avoids potential disruption. Otherwise, switch to Follower normally.
643 RaftActorBehavior behavior = getCurrentBehavior();
644 if (behavior != null && behavior.state() == RaftState.Follower) {
645 String previousLeaderId = behavior.getLeaderId();
646 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
648 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
650 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
652 initializeBehavior();
657 public final DataPersistenceProvider persistence() {
658 return delegatingPersistenceProvider.getDelegate();
661 public void setPersistence(DataPersistenceProvider provider) {
662 delegatingPersistenceProvider.setDelegate(provider);
665 protected void setPersistence(boolean persistent) {
666 DataPersistenceProvider currentPersistence = persistence();
667 if(persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
668 setPersistence(new PersistentDataProvider(this));
670 if(getCurrentBehavior() != null) {
671 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
674 } else if(!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
675 setPersistence(new NonPersistentDataProvider() {
677 * The way snapshotting works is,
679 * <li> RaftActor calls createSnapshot on the Shard
680 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
681 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
682 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
683 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
684 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
688 public void saveSnapshot(Object o) {
689 // Make saving Snapshot successful
690 // Committing the snapshot here would end up calling commit in the creating state which would
691 // be a state violation. That's why now we send a message to commit the snapshot.
692 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
699 * setPeerAddress sets the address of a known peer at a later time.
701 * This is to account for situations where a we know that a peer
702 * exists but we do not know an address up-front. This may also be used in
703 * situations where a known peer starts off in a different location and we
704 * need to change it's address
706 * Note that if the peerId does not match the list of peers passed to
707 * this actor during construction an IllegalStateException will be thrown.
712 protected void setPeerAddress(String peerId, String peerAddress){
713 context.setPeerAddress(peerId, peerAddress);
717 * The applyState method will be called by the RaftActor when some data
718 * needs to be applied to the actor's state
720 * @param clientActor A reference to the client who sent this message. This
721 * is the same reference that was passed to persistData
722 * by the derived actor. clientActor may be null when
723 * the RaftActor is behaving as a follower or during
725 * @param identifier The identifier of the persisted data. This is also
726 * the same identifier that was passed to persistData by
727 * the derived actor. identifier may be null when
728 * the RaftActor is behaving as a follower or during
730 * @param data A piece of data that was persisted by the persistData call.
731 * This should NEVER be null.
733 protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
736 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
739 protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
742 * This method is called when recovery is complete.
744 protected abstract void onRecoveryComplete();
747 * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
750 protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
753 * This method will be called by the RaftActor when the state of the
754 * RaftActor changes. The derived actor can then use methods like
755 * isLeader or getLeader to do something useful
757 protected abstract void onStateChanged();
760 * Notifier Actor for this RaftActor to notify when a role change happens
761 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
763 protected abstract Optional<ActorRef> getRoleChangeNotifier();
766 * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
767 * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
768 * work prior to performing the operation. On completion of any work, the run method must be called on the
769 * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
770 * this actor's thread dispatcher as as it modifies internal state.
772 * The default implementation immediately runs the operation.
774 * @param operation the operation to run
776 protected void pauseLeader(Runnable operation) {
780 protected void onLeaderChanged(String oldLeader, String newLeader) {
784 private String getLeaderAddress(){
786 return getSelf().path().toString();
788 String leaderId = getLeaderId();
789 if (leaderId == null) {
792 String peerAddress = context.getPeerAddress(leaderId);
793 if(LOG.isDebugEnabled()) {
794 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
795 persistenceId(), leaderId, peerAddress);
801 protected boolean hasFollowers(){
802 return getRaftActorContext().hasFollowers();
805 private void captureSnapshot() {
806 SnapshotManager snapshotManager = context.getSnapshotManager();
808 if (!snapshotManager.isCapturing()) {
809 final long idx = getCurrentBehavior().getReplicatedToAllIndex();
810 LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
811 replicatedLog().last(), idx);
813 snapshotManager.capture(replicatedLog().last(), idx);
818 * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
819 * in which case we need to step down.
821 void becomeNonVoting() {
823 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
825 public void onSuccess(ActorRef raftActorRef) {
826 LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
827 ensureFollowerState();
831 public void onFailure(ActorRef raftActorRef) {
832 LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
833 ensureFollowerState();
836 private void ensureFollowerState() {
837 // Whether or not leadership transfer succeeded, we have to step down as leader and
838 // switch to Follower so ensure that.
839 if (getRaftState() != RaftState.Follower) {
840 initializeBehavior();
848 * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.persisted.DeleteEntries}
849 * whose type for fromIndex is long instead of int. This class was kept for backwards
850 * compatibility with Helium.
852 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
853 @SuppressWarnings("serial")
855 static class DeleteEntries implements Serializable {
856 private final int fromIndex;
858 public DeleteEntries(int fromIndex) {
859 this.fromIndex = fromIndex;
862 public int getFromIndex() {
866 private Object readResolve() {
867 return org.opendaylight.controller.cluster.raft.persisted.DeleteEntries.createMigrated(fromIndex);
872 * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm}
873 * which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
875 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
876 @SuppressWarnings("serial")
878 static class UpdateElectionTerm implements Serializable {
879 private final long currentTerm;
880 private final String votedFor;
882 public UpdateElectionTerm(long currentTerm, String votedFor) {
883 this.currentTerm = currentTerm;
884 this.votedFor = votedFor;
887 public long getCurrentTerm() {
891 public String getVotedFor() {
895 private Object readResolve() {
896 return org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm.createMigrated(
897 currentTerm, votedFor);
902 * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
904 private static abstract class BehaviorState implements Immutable {
905 @Nullable abstract RaftActorBehavior getBehavior();
906 @Nullable abstract String getLastValidLeaderId();
907 @Nullable abstract String getLastLeaderId();
908 @Nullable abstract short getLeaderPayloadVersion();
912 * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
914 private static final class SimpleBehaviorState extends BehaviorState {
915 private final RaftActorBehavior behavior;
916 private final String lastValidLeaderId;
917 private final String lastLeaderId;
918 private final short leaderPayloadVersion;
920 SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
921 final RaftActorBehavior behavior) {
922 this.lastValidLeaderId = lastValidLeaderId;
923 this.lastLeaderId = lastLeaderId;
924 this.behavior = Preconditions.checkNotNull(behavior);
925 this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
929 RaftActorBehavior getBehavior() {
934 String getLastValidLeaderId() {
935 return lastValidLeaderId;
939 short getLeaderPayloadVersion() {
940 return leaderPayloadVersion;
944 String getLastLeaderId() {
950 * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
951 * An instance is created for each RaftActor. It has two functions:
952 * - it keeps track of the last leader ID we have encountered since we have been created
953 * - it creates state capture needed to transition from one behavior to the next
955 private static final class BehaviorStateTracker {
957 * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
958 * allowed before we receive the first message, we know the leader ID to be null.
960 private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
962 RaftActorBehavior getBehavior() {
967 String getLastValidLeaderId() {
972 short getLeaderPayloadVersion() {
977 String getLastLeaderId() {
982 private String lastValidLeaderId;
983 private String lastLeaderId;
985 BehaviorState capture(final RaftActorBehavior behavior) {
986 if (behavior == null) {
987 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
988 return NULL_BEHAVIOR_STATE;
991 lastLeaderId = behavior.getLeaderId();
992 if (lastLeaderId != null) {
993 lastValidLeaderId = lastLeaderId;
996 return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);