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.ApplyJournalEntries;
38 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
39 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
40 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
41 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
42 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
44 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
45 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
46 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
48 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
49 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
50 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
51 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
52 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
53 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
54 import org.opendaylight.yangtools.concepts.Identifier;
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 * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
210 * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
211 * implementation for messages which they do not handle
213 * @param message Incoming command message
215 protected void handleNonRaftCommand(final Object message) {
220 * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
221 * {@link #handleNonRaftCommand(Object)} instead.
225 // FIXME: make this method final once our unit tests do not need to override it
226 protected void handleCommand(final Object message) {
227 if (serverConfigurationSupport.handleMessage(message, getSender())) {
230 if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
234 if (message instanceof ApplyState) {
235 ApplyState applyState = (ApplyState) message;
237 long startTime = System.nanoTime();
239 if(LOG.isDebugEnabled()) {
240 LOG.debug("{}: Applying state for log index {} data {}",
241 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
242 applyState.getReplicatedLogEntry().getData());
245 applyState(applyState.getClientActor(), applyState.getIdentifier(),
246 applyState.getReplicatedLogEntry().getData());
248 long elapsedTime = System.nanoTime() - startTime;
249 if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
250 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
251 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
254 if (!hasFollowers()) {
255 // for single node, the capture should happen after the apply state
256 // as we delete messages from the persistent journal which have made it to the snapshot
257 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
258 // and recovery shows data missing
259 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
261 context.getSnapshotManager().trimLog(context.getLastApplied());
264 } else if (message instanceof ApplyJournalEntries) {
265 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
266 if(LOG.isDebugEnabled()) {
267 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
270 persistence().persist(applyEntries, NoopProcedure.instance());
272 } else if (message instanceof FindLeader) {
274 new FindLeaderReply(getLeaderAddress()),
277 } else if(message instanceof GetOnDemandRaftState) {
278 onGetOnDemandRaftStats();
279 } else if(message instanceof InitiateCaptureSnapshot) {
281 } else if(message instanceof SwitchBehavior) {
282 switchBehavior(((SwitchBehavior) message));
283 } else if(message instanceof LeaderTransitioning) {
284 onLeaderTransitioning();
285 } else if(message instanceof Shutdown) {
287 } else if(message instanceof Runnable) {
288 ((Runnable)message).run();
290 // Processing the message may affect the state, hence we need to capture it
291 final RaftActorBehavior currentBehavior = getCurrentBehavior();
292 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
294 // A behavior indicates that it processed the change by returning a reference to the next behavior
295 // to be used. A null return indicates it has not processed the message and we should be passing it to
296 // the subclass for handling.
297 final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
298 if (nextBehavior != null) {
299 switchBehavior(state, nextBehavior);
301 handleNonRaftCommand(message);
306 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
307 LOG.debug("{}: Initiating leader transfer", persistenceId());
309 if(leadershipTransferInProgress == null) {
310 leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
311 leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
313 public void onSuccess(ActorRef raftActorRef) {
314 leadershipTransferInProgress = null;
318 public void onFailure(ActorRef raftActorRef) {
319 leadershipTransferInProgress = null;
323 leadershipTransferInProgress.addOnComplete(onComplete);
324 leadershipTransferInProgress.init();
326 LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
327 leadershipTransferInProgress.addOnComplete(onComplete);
331 private void onShutDown() {
332 LOG.debug("{}: onShutDown", persistenceId());
340 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
341 if (currentBehavior.state() != RaftState.Leader) {
342 // For non-leaders shutdown is a no-op
343 self().tell(PoisonPill.getInstance(), self());
347 if (context.hasFollowers()) {
348 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
350 public void onSuccess(ActorRef raftActorRef) {
351 LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
352 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
356 public void onFailure(ActorRef raftActorRef) {
357 LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
358 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
362 pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
364 protected void doRun() {
365 self().tell(PoisonPill.getInstance(), self());
369 protected void doCancel() {
370 self().tell(PoisonPill.getInstance(), self());
376 private void onLeaderTransitioning() {
377 LOG.debug("{}: onLeaderTransitioning", persistenceId());
378 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
379 if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
380 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
381 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
385 private void switchBehavior(SwitchBehavior message) {
386 if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
387 RaftState newState = message.getNewState();
388 if( newState == RaftState.Leader || newState == RaftState.Follower) {
389 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
390 AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
391 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
393 LOG.warn("Switching to behavior : {} - not supported", newState);
398 private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
399 setCurrentBehavior(nextBehavior);
400 handleBehaviorChange(oldBehaviorState, nextBehavior);
404 RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
405 return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
408 private void onGetOnDemandRaftStats() {
409 // Debugging message to retrieve raft stats.
411 Map<String, String> peerAddresses = new HashMap<>();
412 for(String peerId: context.getPeerIds()) {
413 peerAddresses.put(peerId, context.getPeerAddress(peerId));
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 .peerAddresses(peerAddresses)
433 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
435 ReplicatedLogEntry lastLogEntry = replicatedLog().last();
436 if (lastLogEntry != null) {
437 builder.lastLogIndex(lastLogEntry.getIndex());
438 builder.lastLogTerm(lastLogEntry.getTerm());
441 if(getCurrentBehavior() instanceof AbstractLeader) {
442 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
443 Collection<String> followerIds = leader.getFollowerIds();
444 List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
445 for(String id: followerIds) {
446 final FollowerLogInformation info = leader.getFollower(id);
447 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
448 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity())));
451 builder.followerInfoList(followerInfoList);
454 sender().tell(builder.build(), self());
458 private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
459 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
461 if (oldBehavior != currentBehavior){
465 String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
466 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
468 // it can happen that the state has not changed but the leader has changed.
469 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
470 if(!Objects.equals(lastValidLeaderId, currentBehavior.getLeaderId()) ||
471 oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
472 if(roleChangeNotifier.isPresent()) {
473 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
474 currentBehavior.getLeaderPayloadVersion()), getSelf());
477 onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
479 if(leadershipTransferInProgress != null) {
480 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
483 serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
486 if (roleChangeNotifier.isPresent() &&
487 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
488 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
489 currentBehavior.state().name()), getSelf());
493 protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
494 return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
498 public long snapshotSequenceNr() {
499 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
500 // so that we can delete the persistent journal based on the saved sequence-number
501 // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
502 // was saved and not the number we saved.
503 // We would want to override it , by asking akka to use the last-sequence number known to us.
504 return context.getSnapshotManager().getLastSequenceNumber();
508 * When a derived RaftActor needs to persist something it must call
515 protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
517 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
518 context.getReplicatedLog().lastIndex() + 1,
519 context.getTermInformation().getCurrentTerm(), data);
521 if(LOG.isDebugEnabled()) {
522 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
525 final RaftActorContext raftContext = getRaftActorContext();
527 replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
528 if (!hasFollowers()){
529 // Increment the Commit Index and the Last Applied values
530 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
531 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
533 // Apply the state immediately.
534 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
536 // Send a ApplyJournalEntries message so that we write the fact that we applied
537 // the state to durable storage
538 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
540 } else if (clientActor != null) {
541 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
543 // Send message for replication
544 getCurrentBehavior().handleMessage(getSelf(),
545 new Replicate(clientActor, identifier, replicatedLogEntry1));
550 private ReplicatedLog replicatedLog() {
551 return context.getReplicatedLog();
554 protected String getId() {
555 return context.getId();
559 void setCurrentBehavior(RaftActorBehavior behavior) {
560 context.setCurrentBehavior(behavior);
563 protected RaftActorBehavior getCurrentBehavior() {
564 return context.getCurrentBehavior();
568 * Derived actors can call the isLeader method to check if the current
569 * RaftActor is the Leader or not
571 * @return true it this RaftActor is a Leader false otherwise
573 protected boolean isLeader() {
574 return context.getId().equals(getCurrentBehavior().getLeaderId());
577 protected final boolean isLeaderActive() {
578 return getRaftState() != RaftState.IsolatedLeader && !shuttingDown &&
579 !isLeadershipTransferInProgress();
582 private boolean isLeadershipTransferInProgress() {
583 return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
587 * Derived actor can call getLeader if they need a reference to the Leader.
588 * This would be useful for example in forwarding a request to an actor
589 * which is the leader
591 * @return A reference to the leader if known, null otherwise
593 protected ActorSelection getLeader(){
594 String leaderAddress = getLeaderAddress();
596 if(leaderAddress == null){
600 return context.actorSelection(leaderAddress);
605 * @return the current leader's id
607 protected final String getLeaderId(){
608 return getCurrentBehavior().getLeaderId();
612 protected final RaftState getRaftState() {
613 return getCurrentBehavior().state();
616 protected Long getCurrentTerm(){
617 return context.getTermInformation().getCurrentTerm();
620 protected RaftActorContext getRaftActorContext() {
624 protected void updateConfigParams(ConfigParams configParams) {
626 // obtain the RaftPolicy for oldConfigParams and the updated one.
627 String oldRaftPolicy = context.getConfigParams().
628 getCustomRaftPolicyImplementationClass();
629 String newRaftPolicy = configParams.
630 getCustomRaftPolicyImplementationClass();
632 LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
633 oldRaftPolicy, newRaftPolicy);
634 context.setConfigParams(configParams);
635 if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
636 // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
637 // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
638 // avoids potential disruption. Otherwise, switch to Follower normally.
639 RaftActorBehavior behavior = getCurrentBehavior();
640 if (behavior != null && behavior.state() == RaftState.Follower) {
641 String previousLeaderId = behavior.getLeaderId();
642 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
644 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
646 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
648 initializeBehavior();
653 public final DataPersistenceProvider persistence() {
654 return delegatingPersistenceProvider.getDelegate();
657 public void setPersistence(DataPersistenceProvider provider) {
658 delegatingPersistenceProvider.setDelegate(provider);
661 protected void setPersistence(boolean persistent) {
663 setPersistence(new PersistentDataProvider(this));
665 setPersistence(new NonPersistentDataProvider() {
667 * The way snapshotting works is,
669 * <li> RaftActor calls createSnapshot on the Shard
670 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
671 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
672 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
673 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
674 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
678 public void saveSnapshot(Object o) {
679 // Make saving Snapshot successful
680 // Committing the snapshot here would end up calling commit in the creating state which would
681 // be a state violation. That's why now we send a message to commit the snapshot.
682 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
689 * setPeerAddress sets the address of a known peer at a later time.
691 * This is to account for situations where a we know that a peer
692 * exists but we do not know an address up-front. This may also be used in
693 * situations where a known peer starts off in a different location and we
694 * need to change it's address
696 * Note that if the peerId does not match the list of peers passed to
697 * this actor during construction an IllegalStateException will be thrown.
702 protected void setPeerAddress(String peerId, String peerAddress){
703 context.setPeerAddress(peerId, peerAddress);
707 * The applyState method will be called by the RaftActor when some data
708 * needs to be applied to the actor's state
710 * @param clientActor A reference to the client who sent this message. This
711 * is the same reference that was passed to persistData
712 * by the derived actor. clientActor may be null when
713 * the RaftActor is behaving as a follower or during
715 * @param identifier The identifier of the persisted data. This is also
716 * the same identifier that was passed to persistData by
717 * the derived actor. identifier may be null when
718 * the RaftActor is behaving as a follower or during
720 * @param data A piece of data that was persisted by the persistData call.
721 * This should NEVER be null.
723 protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
726 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
729 protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
732 * This method is called when recovery is complete.
734 protected abstract void onRecoveryComplete();
737 * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
740 protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
743 * This method will be called by the RaftActor when the state of the
744 * RaftActor changes. The derived actor can then use methods like
745 * isLeader or getLeader to do something useful
747 protected abstract void onStateChanged();
750 * Notifier Actor for this RaftActor to notify when a role change happens
751 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
753 protected abstract Optional<ActorRef> getRoleChangeNotifier();
756 * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
757 * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
758 * work prior to performing the operation. On completion of any work, the run method must be called on the
759 * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
760 * this actor's thread dispatcher as as it modifies internal state.
762 * The default implementation immediately runs the operation.
764 * @param operation the operation to run
766 protected void pauseLeader(Runnable operation) {
770 protected void onLeaderChanged(String oldLeader, String newLeader) {
774 private String getLeaderAddress(){
776 return getSelf().path().toString();
778 String leaderId = getLeaderId();
779 if (leaderId == null) {
782 String peerAddress = context.getPeerAddress(leaderId);
783 if(LOG.isDebugEnabled()) {
784 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
785 persistenceId(), leaderId, peerAddress);
791 protected boolean hasFollowers(){
792 return getRaftActorContext().hasFollowers();
795 private void captureSnapshot() {
796 SnapshotManager snapshotManager = context.getSnapshotManager();
798 if (!snapshotManager.isCapturing()) {
799 final long idx = getCurrentBehavior().getReplicatedToAllIndex();
800 LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
801 replicatedLog().last(), idx);
803 snapshotManager.capture(replicatedLog().last(), idx);
808 * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
809 * in which case we need to step down.
811 void becomeNonVoting() {
813 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
815 public void onSuccess(ActorRef raftActorRef) {
816 LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
817 ensureFollowerState();
821 public void onFailure(ActorRef raftActorRef) {
822 LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
823 ensureFollowerState();
826 private void ensureFollowerState() {
827 // Whether or not leadership transfer succeeded, we have to step down as leader and
828 // switch to Follower so ensure that.
829 if (getRaftState() != RaftState.Follower) {
830 initializeBehavior();
838 * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
839 * whose type for fromIndex is long instead of int. This class was kept for backwards
840 * compatibility with Helium.
842 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
843 @SuppressWarnings("serial")
845 static class DeleteEntries implements Serializable {
846 private final int fromIndex;
848 public DeleteEntries(int fromIndex) {
849 this.fromIndex = fromIndex;
852 public int getFromIndex() {
858 * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
859 * which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
861 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
862 @SuppressWarnings("serial")
864 static class UpdateElectionTerm implements Serializable {
865 private final long currentTerm;
866 private final String votedFor;
868 public UpdateElectionTerm(long currentTerm, String votedFor) {
869 this.currentTerm = currentTerm;
870 this.votedFor = votedFor;
873 public long getCurrentTerm() {
877 public String getVotedFor() {
883 * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
885 private static abstract class BehaviorState implements Immutable {
886 @Nullable abstract RaftActorBehavior getBehavior();
887 @Nullable abstract String getLastValidLeaderId();
888 @Nullable abstract short getLeaderPayloadVersion();
892 * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
894 private static final class SimpleBehaviorState extends BehaviorState {
895 private final RaftActorBehavior behavior;
896 private final String lastValidLeaderId;
897 private final short leaderPayloadVersion;
899 SimpleBehaviorState(final String lastValidLeaderId, final RaftActorBehavior behavior) {
900 this.lastValidLeaderId = lastValidLeaderId;
901 this.behavior = Preconditions.checkNotNull(behavior);
902 this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
906 RaftActorBehavior getBehavior() {
911 String getLastValidLeaderId() {
912 return lastValidLeaderId;
916 short getLeaderPayloadVersion() {
917 return leaderPayloadVersion;
922 * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
923 * An instance is created for each RaftActor. It has two functions:
924 * - it keeps track of the last leader ID we have encountered since we have been created
925 * - it creates state capture needed to transition from one behavior to the next
927 private static final class BehaviorStateTracker {
929 * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
930 * allowed before we receive the first message, we know the leader ID to be null.
932 private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
934 RaftActorBehavior getBehavior() {
939 String getLastValidLeaderId() {
944 short getLeaderPayloadVersion() {
949 private String lastValidLeaderId;
951 BehaviorState capture(final RaftActorBehavior behavior) {
952 if (behavior == null) {
953 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
954 return NULL_BEHAVIOR_STATE;
957 final String leaderId = behavior.getLeaderId();
958 if (leaderId != null) {
959 lastValidLeaderId = leaderId;
962 return new SimpleBehaviorState(lastValidLeaderId, behavior);