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.util.Collection;
21 import java.util.HashMap;
22 import java.util.List;
24 import java.util.Objects;
25 import java.util.concurrent.TimeUnit;
26 import javax.annotation.Nonnull;
27 import javax.annotation.Nullable;
28 import org.apache.commons.lang3.time.DurationFormatUtils;
29 import org.opendaylight.controller.cluster.DataPersistenceProvider;
30 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
31 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
32 import org.opendaylight.controller.cluster.PersistentDataProvider;
33 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
34 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
35 import org.opendaylight.controller.cluster.notifications.RoleChanged;
36 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
37 import org.opendaylight.controller.cluster.raft.base.messages.CheckConsensusReached;
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;
59 * RaftActor encapsulates a state machine that needs to be kept synchronized
60 * in a cluster. It implements the RAFT algorithm as described in the paper
61 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
62 * 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
93 * <li> when a log entry should be persisted </li>
94 * <li> when a log entry should be applied to the state machine (and) </li>
95 * <li> when a snapshot should be saved </li>
98 public abstract class RaftActor extends AbstractUntypedPersistentActor {
100 private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
103 * This context should NOT be passed directly to any other actor it is
104 * only to be consumed by the RaftActorBehaviors.
106 private final RaftActorContextImpl context;
108 private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
110 private final PersistentDataProvider persistentProvider;
112 private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
114 private RaftActorRecoverySupport raftRecovery;
116 private RaftActorSnapshotMessageSupport snapshotSupport;
118 private RaftActorServerConfigurationSupport serverConfigurationSupport;
120 private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
122 private boolean shuttingDown;
124 protected RaftActor(String id, Map<String, String> peerAddresses,
125 Optional<ConfigParams> configParams, short payloadVersion) {
127 persistentProvider = new PersistentDataProvider(this);
128 delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
130 context = new RaftActorContextImpl(this.getSelf(),
131 this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
132 -1, -1, peerAddresses,
133 configParams.isPresent() ? configParams.get() : new DefaultConfigParamsImpl(),
134 delegatingPersistenceProvider, LOG);
136 context.setPayloadVersion(payloadVersion);
137 context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
141 public void preStart() throws Exception {
142 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
143 context.getConfigParams().getJournalRecoveryLogBatchSize());
147 snapshotSupport = newRaftActorSnapshotMessageSupport();
148 serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
152 public void postStop() {
158 protected void handleRecover(Object message) {
159 if (raftRecovery == null) {
160 raftRecovery = newRaftActorRecoverySupport();
163 boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
164 if (recoveryComplete) {
165 onRecoveryComplete();
167 initializeBehavior();
173 protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
174 return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
178 void initializeBehavior() {
179 changeCurrentBehavior(new Follower(context));
183 @SuppressWarnings("checkstyle:IllegalCatch")
184 protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
185 final RaftActorBehavior currentBehavior = getCurrentBehavior();
186 if (currentBehavior != null) {
188 currentBehavior.close();
189 } catch (Exception e) {
190 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
194 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
195 setCurrentBehavior(newBehavior);
196 handleBehaviorChange(state, newBehavior);
200 * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
201 * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
202 * implementation for messages which they do not handle
204 * @param message Incoming command message
206 protected void handleNonRaftCommand(final Object message) {
213 * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
214 * {@link #handleNonRaftCommand(Object)} instead.
218 // FIXME: make this method final once our unit tests do not need to override it
219 protected void handleCommand(final Object message) {
220 if (serverConfigurationSupport.handleMessage(message, getSender())) {
223 if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
227 if (message instanceof ApplyState) {
228 ApplyState applyState = (ApplyState) message;
230 long startTime = System.nanoTime();
232 if (LOG.isDebugEnabled()) {
233 LOG.debug("{}: Applying state for log index {} data {}",
234 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
235 applyState.getReplicatedLogEntry().getData());
238 if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
239 applyState(applyState.getClientActor(), applyState.getIdentifier(),
240 applyState.getReplicatedLogEntry().getData());
243 long elapsedTime = System.nanoTime() - startTime;
244 if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
245 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
246 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
249 if (!hasFollowers()) {
250 // for single node, the capture should happen after the apply state
251 // as we delete messages from the persistent journal which have made it to the snapshot
252 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
253 // and recovery shows data missing
254 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
256 context.getSnapshotManager().trimLog(context.getLastApplied());
259 // Send it to the current behavior - some behaviors like PreLeader need to be notified of ApplyState.
260 possiblyHandleBehaviorMessage(message);
262 } else if (message instanceof ApplyJournalEntries) {
263 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
264 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
266 persistence().persistAsync(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);
287 } else if (!possiblyHandleBehaviorMessage(message)) {
288 handleNonRaftCommand(message);
292 private boolean possiblyHandleBehaviorMessage(final Object message) {
293 final RaftActorBehavior currentBehavior = getCurrentBehavior();
294 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
296 // A behavior indicates that it processed the change by returning a reference to the next behavior
297 // to be used. A null return indicates it has not processed the message and we should be passing it to
298 // the subclass for handling.
299 final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
300 if (nextBehavior != null) {
301 switchBehavior(state, nextBehavior);
308 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
309 LOG.debug("{}: Initiating leader transfer", persistenceId());
311 if (leadershipTransferInProgress == null) {
312 leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
313 leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
315 public void onSuccess(ActorRef raftActorRef) {
316 leadershipTransferInProgress = null;
320 public void onFailure(ActorRef raftActorRef) {
321 leadershipTransferInProgress = null;
325 leadershipTransferInProgress.addOnComplete(onComplete);
326 leadershipTransferInProgress.init();
328 LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
329 leadershipTransferInProgress.addOnComplete(onComplete);
333 private void onShutDown() {
334 LOG.debug("{}: onShutDown", persistenceId());
342 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
343 if (currentBehavior.state() != RaftState.Leader) {
344 // For non-leaders shutdown is a no-op
345 self().tell(PoisonPill.getInstance(), self());
349 if (context.hasFollowers()) {
350 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
352 public void onSuccess(ActorRef raftActorRef) {
353 LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
354 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
358 public void onFailure(ActorRef raftActorRef) {
359 LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
360 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
364 pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
366 protected void doRun() {
367 self().tell(PoisonPill.getInstance(), self());
371 protected void doCancel() {
372 self().tell(PoisonPill.getInstance(), self());
378 private void onLeaderTransitioning() {
379 LOG.debug("{}: onLeaderTransitioning", persistenceId());
380 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
381 if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
382 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
383 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
387 private void switchBehavior(SwitchBehavior message) {
388 if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
389 RaftState newState = message.getNewState();
390 if ( newState == RaftState.Leader || newState == RaftState.Follower) {
391 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
392 AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
393 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
395 LOG.warn("Switching to behavior : {} - not supported", newState);
400 private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
401 setCurrentBehavior(nextBehavior);
402 handleBehaviorChange(oldBehaviorState, nextBehavior);
406 RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
407 return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
410 private void onGetOnDemandRaftStats() {
411 // Debugging message to retrieve raft stats.
413 Map<String, String> peerAddresses = new HashMap<>();
414 Map<String, Boolean> peerVotingStates = new HashMap<>();
415 for (PeerInfo info: context.getPeers()) {
416 peerVotingStates.put(info.getId(), info.isVoting());
417 peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
420 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
421 OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
422 .commitIndex(context.getCommitIndex())
423 .currentTerm(context.getTermInformation().getCurrentTerm())
424 .inMemoryJournalDataSize(replicatedLog().dataSize())
425 .inMemoryJournalLogSize(replicatedLog().size())
426 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
427 .lastApplied(context.getLastApplied())
428 .lastIndex(replicatedLog().lastIndex())
429 .lastTerm(replicatedLog().lastTerm())
430 .leader(getLeaderId())
431 .raftState(currentBehavior.state().toString())
432 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
433 .snapshotIndex(replicatedLog().getSnapshotIndex())
434 .snapshotTerm(replicatedLog().getSnapshotTerm())
435 .votedFor(context.getTermInformation().getVotedFor())
436 .isVoting(context.isVotingMember())
437 .peerAddresses(peerAddresses)
438 .peerVotingStates(peerVotingStates)
439 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
441 ReplicatedLogEntry lastLogEntry = replicatedLog().last();
442 if (lastLogEntry != null) {
443 builder.lastLogIndex(lastLogEntry.getIndex());
444 builder.lastLogTerm(lastLogEntry.getTerm());
447 if (getCurrentBehavior() instanceof AbstractLeader) {
448 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
449 Collection<String> followerIds = leader.getFollowerIds();
450 List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
451 for (String id: followerIds) {
452 final FollowerLogInformation info = leader.getFollower(id);
453 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
454 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
455 context.getPeerInfo(info.getId()).isVoting()));
458 builder.followerInfoList(followerInfoList);
461 sender().tell(builder.build(), self());
465 private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
466 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
468 if (oldBehavior != currentBehavior) {
472 String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
473 String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
474 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
476 // it can happen that the state has not changed but the leader has changed.
477 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
478 if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
479 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
480 if (roleChangeNotifier.isPresent()) {
481 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
482 currentBehavior.getLeaderPayloadVersion()), getSelf());
485 onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
487 if (leadershipTransferInProgress != null) {
488 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
491 serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
494 if (roleChangeNotifier.isPresent()
495 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
496 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
497 currentBehavior.state().name()), getSelf());
501 protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
502 return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
506 public long snapshotSequenceNr() {
507 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
508 // so that we can delete the persistent journal based on the saved sequence-number
509 // However , when akka replays the journal during recovery, it replays it from the sequence number when the
510 // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
511 // last-sequence number known to us.
512 return context.getSnapshotManager().getLastSequenceNumber();
516 * 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);
524 replicatedLogEntry.setPersistencePending(true);
526 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
528 final RaftActorContext raftContext = getRaftActorContext();
530 boolean wasAppended = replicatedLog().appendAndPersist(replicatedLogEntry, persistedLogEntry -> {
531 // Clear the persistence pending flag in the log entry.
532 persistedLogEntry.setPersistencePending(false);
534 if (!hasFollowers()) {
535 // Increment the Commit Index and the Last Applied values
536 raftContext.setCommitIndex(persistedLogEntry.getIndex());
537 raftContext.setLastApplied(persistedLogEntry.getIndex());
539 // Apply the state immediately.
540 self().tell(new ApplyState(clientActor, identifier, persistedLogEntry), self());
542 // Send a ApplyJournalEntries message so that we write the fact that we applied
543 // the state to durable storage
544 self().tell(new ApplyJournalEntries(persistedLogEntry.getIndex()), self());
547 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
549 // Local persistence is complete so send the CheckConsensusReached message to the behavior (which
550 // normally should still be the leader) to check if consensus has now been reached in conjunction with
551 // follower replication.
552 getCurrentBehavior().handleMessage(getSelf(), CheckConsensusReached.INSTANCE);
556 if (wasAppended && hasFollowers()) {
557 // Send log entry for replication.
558 getCurrentBehavior().handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry));
562 private ReplicatedLog replicatedLog() {
563 return context.getReplicatedLog();
566 protected String getId() {
567 return context.getId();
571 void setCurrentBehavior(RaftActorBehavior behavior) {
572 context.setCurrentBehavior(behavior);
575 protected RaftActorBehavior getCurrentBehavior() {
576 return context.getCurrentBehavior();
580 * Derived actors can call the isLeader method to check if the current
581 * RaftActor is the Leader or not.
583 * @return true it this RaftActor is a Leader false otherwise
585 protected boolean isLeader() {
586 return context.getId().equals(getCurrentBehavior().getLeaderId());
589 protected final boolean isLeaderActive() {
590 return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
591 && !shuttingDown && !isLeadershipTransferInProgress();
594 private boolean isLeadershipTransferInProgress() {
595 return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
599 * Derived actor can call getLeader if they need a reference to the Leader.
600 * This would be useful for example in forwarding a request to an actor
601 * which is the leader
603 * @return A reference to the leader if known, null otherwise
605 public ActorSelection getLeader() {
606 String leaderAddress = getLeaderAddress();
608 if (leaderAddress == null) {
612 return context.actorSelection(leaderAddress);
616 * Returns the id of the current leader.
618 * @return the current leader's id
620 protected final String getLeaderId() {
621 return getCurrentBehavior().getLeaderId();
625 protected final RaftState getRaftState() {
626 return getCurrentBehavior().state();
629 protected Long getCurrentTerm() {
630 return context.getTermInformation().getCurrentTerm();
633 protected RaftActorContext getRaftActorContext() {
637 protected void updateConfigParams(ConfigParams configParams) {
639 // obtain the RaftPolicy for oldConfigParams and the updated one.
640 String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
641 String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
643 LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
644 oldRaftPolicy, newRaftPolicy);
645 context.setConfigParams(configParams);
646 if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
647 // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
648 // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
649 // avoids potential disruption. Otherwise, switch to Follower normally.
650 RaftActorBehavior behavior = getCurrentBehavior();
651 if (behavior != null && behavior.state() == RaftState.Follower) {
652 String previousLeaderId = behavior.getLeaderId();
653 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
655 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
658 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
660 initializeBehavior();
665 public final DataPersistenceProvider persistence() {
666 return delegatingPersistenceProvider.getDelegate();
669 public void setPersistence(DataPersistenceProvider provider) {
670 delegatingPersistenceProvider.setDelegate(provider);
673 protected void setPersistence(boolean persistent) {
674 DataPersistenceProvider currentPersistence = persistence();
675 if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
676 setPersistence(new PersistentDataProvider(this));
678 if (getCurrentBehavior() != null) {
679 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
682 } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
683 setPersistence(new NonPersistentDataProvider() {
685 * The way snapshotting works is,
687 * <li> RaftActor calls createSnapshot on the Shard
688 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
689 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
690 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
691 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
692 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
696 public void saveSnapshot(Object object) {
697 // Make saving Snapshot successful
698 // Committing the snapshot here would end up calling commit in the creating state which would
699 // be a state violation. That's why now we send a message to commit the snapshot.
700 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
707 * setPeerAddress sets the address of a known peer at a later time.
710 * This is to account for situations where a we know that a peer
711 * exists but we do not know an address up-front. This may also be used in
712 * situations where a known peer starts off in a different location and we
713 * need to change it's address
716 * Note that if the peerId does not match the list of peers passed to
717 * this actor during construction an IllegalStateException will be thrown.
719 protected void setPeerAddress(String peerId, String peerAddress) {
720 context.setPeerAddress(peerId, peerAddress);
724 * The applyState method will be called by the RaftActor when some data
725 * needs to be applied to the actor's state.
727 * @param clientActor A reference to the client who sent this message. This
728 * is the same reference that was passed to persistData
729 * by the derived actor. clientActor may be null when
730 * the RaftActor is behaving as a follower or during
732 * @param identifier The identifier of the persisted data. This is also
733 * the same identifier that was passed to persistData by
734 * the derived actor. identifier may be null when
735 * the RaftActor is behaving as a follower or during
737 * @param data A piece of data that was persisted by the persistData call.
738 * This should NEVER be null.
740 protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
743 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
746 protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
749 * This method is called when recovery is complete.
751 protected abstract void onRecoveryComplete();
754 * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
757 protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
760 * This method will be called by the RaftActor when the state of the
761 * RaftActor changes. The derived actor can then use methods like
762 * isLeader or getLeader to do something useful
764 protected abstract void onStateChanged();
767 * Notifier Actor for this RaftActor to notify when a role change happens.
769 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
771 protected abstract Optional<ActorRef> getRoleChangeNotifier();
774 * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
775 * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
776 * work prior to performing the operation. On completion of any work, the run method must be called on the
777 * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
778 * this actor's thread dispatcher as as it modifies internal state.
781 * The default implementation immediately runs the operation.
783 * @param operation the operation to run
785 protected void pauseLeader(Runnable operation) {
789 protected void onLeaderChanged(String oldLeader, String newLeader) {
792 private String getLeaderAddress() {
794 return getSelf().path().toString();
796 String leaderId = getLeaderId();
797 if (leaderId == null) {
800 String peerAddress = context.getPeerAddress(leaderId);
801 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
806 protected boolean hasFollowers() {
807 return getRaftActorContext().hasFollowers();
810 private void captureSnapshot() {
811 SnapshotManager snapshotManager = context.getSnapshotManager();
813 if (!snapshotManager.isCapturing()) {
814 final long idx = getCurrentBehavior().getReplicatedToAllIndex();
815 LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
816 replicatedLog().last(), idx);
818 snapshotManager.capture(replicatedLog().last(), idx);
823 * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
824 * in which case we need to step down.
826 void becomeNonVoting() {
828 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
830 public void onSuccess(ActorRef raftActorRef) {
831 LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
832 ensureFollowerState();
836 public void onFailure(ActorRef raftActorRef) {
837 LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
838 ensureFollowerState();
841 private void ensureFollowerState() {
842 // Whether or not leadership transfer succeeded, we have to step down as leader and
843 // switch to Follower so ensure that.
844 if (getRaftState() != RaftState.Follower) {
845 initializeBehavior();
853 * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
855 private abstract static class BehaviorState implements Immutable {
856 @Nullable abstract RaftActorBehavior getBehavior();
858 @Nullable abstract String getLastValidLeaderId();
860 @Nullable abstract String getLastLeaderId();
862 @Nullable abstract short getLeaderPayloadVersion();
866 * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
868 private static final class SimpleBehaviorState extends BehaviorState {
869 private final RaftActorBehavior behavior;
870 private final String lastValidLeaderId;
871 private final String lastLeaderId;
872 private final short leaderPayloadVersion;
874 SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
875 final RaftActorBehavior behavior) {
876 this.lastValidLeaderId = lastValidLeaderId;
877 this.lastLeaderId = lastLeaderId;
878 this.behavior = Preconditions.checkNotNull(behavior);
879 this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
883 RaftActorBehavior getBehavior() {
888 String getLastValidLeaderId() {
889 return lastValidLeaderId;
893 short getLeaderPayloadVersion() {
894 return leaderPayloadVersion;
898 String getLastLeaderId() {
904 * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
905 * An instance is created for each RaftActor. It has two functions:
906 * - it keeps track of the last leader ID we have encountered since we have been created
907 * - it creates state capture needed to transition from one behavior to the next
909 private static final class BehaviorStateTracker {
911 * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
912 * allowed before we receive the first message, we know the leader ID to be null.
914 private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
916 RaftActorBehavior getBehavior() {
921 String getLastValidLeaderId() {
926 short getLeaderPayloadVersion() {
931 String getLastLeaderId() {
936 private String lastValidLeaderId;
937 private String lastLeaderId;
939 BehaviorState capture(final RaftActorBehavior behavior) {
940 if (behavior == null) {
941 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
942 return NULL_BEHAVIOR_STATE;
945 lastLeaderId = behavior.getLeaderId();
946 if (lastLeaderId != null) {
947 lastValidLeaderId = lastLeaderId;
950 return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);