2 * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved.
3 * Copyright (c) 2015 Brocade Communications Systems, Inc. and others. All rights reserved.
5 * This program and the accompanying materials are made available under the
6 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
7 * and is available at http://www.eclipse.org/legal/epl-v10.html
10 package org.opendaylight.controller.cluster.raft;
12 import akka.actor.ActorRef;
13 import akka.actor.ActorSelection;
14 import akka.actor.PoisonPill;
15 import akka.actor.Status;
16 import com.google.common.annotations.VisibleForTesting;
17 import com.google.common.base.Optional;
18 import com.google.common.base.Preconditions;
19 import com.google.common.base.Verify;
20 import com.google.common.collect.Lists;
21 import java.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.CheckConsensusReached;
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.messages.RequestLeadership;
54 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
55 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
56 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
57 import org.opendaylight.controller.cluster.raft.persisted.SimpleReplicatedLogEntry;
58 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
59 import org.opendaylight.yangtools.concepts.Identifier;
60 import org.opendaylight.yangtools.concepts.Immutable;
63 * RaftActor encapsulates a state machine that needs to be kept synchronized
64 * in a cluster. It implements the RAFT algorithm as described in the paper
65 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
66 * In Search of an Understandable Consensus Algorithm</a>
69 * RaftActor has 3 states and each state has a certain behavior associated
70 * with it. A Raft actor can behave as,
73 * <li> A Follower (or) </li>
74 * <li> A Candidate </li>
78 * A RaftActor MUST be a Leader in order to accept requests from clients to
79 * change the state of it's encapsulated state machine. Once a RaftActor becomes
80 * a Leader it is also responsible for ensuring that all followers ultimately
81 * have the same log and therefore the same state machine as itself.
84 * The current behavior of a RaftActor determines how election for leadership
85 * is initiated and how peer RaftActors react to request for votes.
88 * Each RaftActor also needs to know the current election term. It uses this
89 * information for a couple of things. One is to simply figure out who it
90 * voted for in the last election. Another is to figure out if the message
91 * it received to update it's state is stale.
94 * The RaftActor uses akka-persistence to store it's replicated log.
95 * Furthermore through it's behaviors a Raft Actor determines
97 * <li> when a log entry should be persisted </li>
98 * <li> when a log entry should be applied to the state machine (and) </li>
99 * <li> when a snapshot should be saved </li>
102 public abstract class RaftActor extends AbstractUntypedPersistentActor {
104 private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
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 protected 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, this::handleApplyState, 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 @SuppressWarnings("checkstyle:IllegalCatch")
188 protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
189 final RaftActorBehavior currentBehavior = getCurrentBehavior();
190 if (currentBehavior != null) {
192 currentBehavior.close();
193 } catch (Exception e) {
194 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
198 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
199 setCurrentBehavior(newBehavior);
200 handleBehaviorChange(state, newBehavior);
204 * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
205 * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
206 * implementation for messages which they do not handle
208 * @param message Incoming command message
210 protected void handleNonRaftCommand(final Object message) {
217 * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
218 * {@link #handleNonRaftCommand(Object)} instead.
222 // FIXME: make this method final once our unit tests do not need to override it
223 protected void handleCommand(final Object message) {
224 if (serverConfigurationSupport.handleMessage(message, getSender())) {
227 if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
230 if (message instanceof ApplyState) {
231 ApplyState applyState = (ApplyState) message;
233 if (!hasFollowers()) {
234 // for single node, the capture should happen after the apply state
235 // as we delete messages from the persistent journal which have made it to the snapshot
236 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
237 // and recovery shows data missing
238 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
240 context.getSnapshotManager().trimLog(context.getLastApplied());
243 possiblyHandleBehaviorMessage(message);
244 } else if (message instanceof ApplyJournalEntries) {
245 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
246 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
248 persistence().persistAsync(applyEntries, NoopProcedure.instance());
250 } else if (message instanceof FindLeader) {
252 new FindLeaderReply(getLeaderAddress()),
255 } else if (message instanceof GetOnDemandRaftState) {
256 onGetOnDemandRaftStats();
257 } else if (message instanceof InitiateCaptureSnapshot) {
259 } else if (message instanceof SwitchBehavior) {
260 switchBehavior((SwitchBehavior) message);
261 } else if (message instanceof LeaderTransitioning) {
262 onLeaderTransitioning((LeaderTransitioning)message);
263 } else if (message instanceof Shutdown) {
265 } else if (message instanceof Runnable) {
266 ((Runnable)message).run();
267 } else if (message instanceof NoopPayload) {
268 persistData(null, null, (NoopPayload) message, false);
269 } else if (message instanceof RequestLeadership) {
270 onRequestLeadership((RequestLeadership) message);
271 } else if (!possiblyHandleBehaviorMessage(message)) {
272 handleNonRaftCommand(message);
276 private void onRequestLeadership(final RequestLeadership message) {
277 LOG.debug("{}: onRequestLeadership {}", persistenceId(), message);
279 // non-leader cannot satisfy leadership request
280 LOG.warn("{}: onRequestLeadership {} was sent to non-leader."
281 + " Current behavior: {}. Sending failure response",
282 persistenceId(), getCurrentBehavior().state());
283 message.getReplyTo().tell(new LeadershipTransferFailedException("Cannot transfer leader to "
284 + message.getRequestedFollowerId()
285 + ". RequestLeadership message was sent to non-leader " + persistenceId()), getSelf());
289 final String requestedFollowerId = message.getRequestedFollowerId();
290 final ActorRef replyTo = message.getReplyTo();
291 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
293 public void onSuccess(final ActorRef raftActorRef) {
295 if (!requestedFollowerId.equals(getLeaderId())) {
296 onFailure(raftActorRef);
299 LOG.debug("{}: Leadership transferred successfully to {}", persistenceId(), requestedFollowerId);
300 replyTo.tell(new Status.Success(null), getSelf());
304 public void onFailure(final ActorRef raftActorRef) {
305 LOG.debug("{}: LeadershipTransfer request from {} failed", persistenceId(), requestedFollowerId);
306 replyTo.tell(new Status.Failure(
307 new LeadershipTransferFailedException(
308 "Failed to transfer leadership to " + requestedFollowerId
309 + ". Follower is not ready to become leader")),
312 }, message.getRequestedFollowerId());
315 private boolean possiblyHandleBehaviorMessage(final Object message) {
316 final RaftActorBehavior currentBehavior = getCurrentBehavior();
317 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
319 // A behavior indicates that it processed the change by returning a reference to the next behavior
320 // to be used. A null return indicates it has not processed the message and we should be passing it to
321 // the subclass for handling.
322 final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
323 if (nextBehavior != null) {
324 switchBehavior(state, nextBehavior);
331 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
332 initiateLeadershipTransfer(onComplete, null);
335 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete,
336 final String followerId) {
337 LOG.debug("{}: Initiating leader transfer", persistenceId());
339 if (leadershipTransferInProgress == null) {
340 leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this, followerId);
341 leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
343 public void onSuccess(ActorRef raftActorRef) {
344 leadershipTransferInProgress = null;
348 public void onFailure(ActorRef raftActorRef) {
349 leadershipTransferInProgress = null;
353 leadershipTransferInProgress.addOnComplete(onComplete);
354 leadershipTransferInProgress.init();
356 LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
357 leadershipTransferInProgress.addOnComplete(onComplete);
361 private void onShutDown() {
362 LOG.debug("{}: onShutDown", persistenceId());
370 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
371 if (currentBehavior.state() != RaftState.Leader) {
372 // For non-leaders shutdown is a no-op
373 self().tell(PoisonPill.getInstance(), self());
377 if (context.hasFollowers()) {
378 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
380 public void onSuccess(ActorRef raftActorRef) {
381 LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
382 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
386 public void onFailure(ActorRef raftActorRef) {
387 LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
388 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
392 pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
394 protected void doRun() {
395 self().tell(PoisonPill.getInstance(), self());
399 protected void doCancel() {
400 self().tell(PoisonPill.getInstance(), self());
406 private void onLeaderTransitioning(final LeaderTransitioning leaderTransitioning) {
407 LOG.debug("{}: onLeaderTransitioning: {}", persistenceId(), leaderTransitioning);
408 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
409 if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()
410 && leaderTransitioning.getLeaderId().equals(getCurrentBehavior().getLeaderId())) {
411 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
412 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
416 private void switchBehavior(SwitchBehavior message) {
417 if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
418 RaftState newState = message.getNewState();
419 if (newState == RaftState.Leader || newState == RaftState.Follower) {
420 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
421 AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
422 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
424 LOG.warn("Switching to behavior : {} - not supported", newState);
429 private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
430 setCurrentBehavior(nextBehavior);
431 handleBehaviorChange(oldBehaviorState, nextBehavior);
435 RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
436 return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
439 private void onGetOnDemandRaftStats() {
440 // Debugging message to retrieve raft stats.
442 Map<String, String> peerAddresses = new HashMap<>();
443 Map<String, Boolean> peerVotingStates = new HashMap<>();
444 for (PeerInfo info: context.getPeers()) {
445 peerVotingStates.put(info.getId(), info.isVoting());
446 peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
449 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
450 OnDemandRaftState.AbstractBuilder<?> builder = newOnDemandRaftStateBuilder()
451 .commitIndex(context.getCommitIndex())
452 .currentTerm(context.getTermInformation().getCurrentTerm())
453 .inMemoryJournalDataSize(replicatedLog().dataSize())
454 .inMemoryJournalLogSize(replicatedLog().size())
455 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
456 .lastApplied(context.getLastApplied())
457 .lastIndex(replicatedLog().lastIndex())
458 .lastTerm(replicatedLog().lastTerm())
459 .leader(getLeaderId())
460 .raftState(currentBehavior.state().toString())
461 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
462 .snapshotIndex(replicatedLog().getSnapshotIndex())
463 .snapshotTerm(replicatedLog().getSnapshotTerm())
464 .votedFor(context.getTermInformation().getVotedFor())
465 .isVoting(context.isVotingMember())
466 .peerAddresses(peerAddresses)
467 .peerVotingStates(peerVotingStates)
468 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
470 ReplicatedLogEntry lastLogEntry = replicatedLog().last();
471 if (lastLogEntry != null) {
472 builder.lastLogIndex(lastLogEntry.getIndex());
473 builder.lastLogTerm(lastLogEntry.getTerm());
476 if (getCurrentBehavior() instanceof AbstractLeader) {
477 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
478 Collection<String> followerIds = leader.getFollowerIds();
479 List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
480 for (String id: followerIds) {
481 final FollowerLogInformation info = leader.getFollower(id);
482 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
483 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
484 context.getPeerInfo(info.getId()).isVoting()));
487 builder.followerInfoList(followerInfoList);
490 sender().tell(builder.build(), self());
494 protected OnDemandRaftState.AbstractBuilder<?> newOnDemandRaftStateBuilder() {
495 return OnDemandRaftState.builder();
498 private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
499 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
501 if (oldBehavior != currentBehavior) {
505 String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
506 String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
507 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
509 // it can happen that the state has not changed but the leader has changed.
510 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
511 if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
512 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
513 if (roleChangeNotifier.isPresent()) {
514 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
515 currentBehavior.getLeaderPayloadVersion()), getSelf());
518 onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
520 if (leadershipTransferInProgress != null) {
521 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
524 serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
527 if (roleChangeNotifier.isPresent()
528 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
529 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
530 currentBehavior.state().name()), getSelf());
534 private void handleApplyState(ApplyState applyState) {
535 long startTime = System.nanoTime();
537 Payload payload = applyState.getReplicatedLogEntry().getData();
538 if (LOG.isDebugEnabled()) {
539 LOG.debug("{}: Applying state for log index {} data {}",
540 persistenceId(), applyState.getReplicatedLogEntry().getIndex(), payload);
543 if (!(payload instanceof NoopPayload) && !(payload instanceof ServerConfigurationPayload)) {
544 applyState(applyState.getClientActor(), applyState.getIdentifier(), payload);
547 long elapsedTime = System.nanoTime() - startTime;
548 if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
549 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
550 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
553 // Send the ApplyState message back to self to handle further processing asynchronously.
554 self().tell(applyState, self());
557 protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
558 return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
562 public long snapshotSequenceNr() {
563 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
564 // so that we can delete the persistent journal based on the saved sequence-number
565 // However , when akka replays the journal during recovery, it replays it from the sequence number when the
566 // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
567 // last-sequence number known to us.
568 return context.getSnapshotManager().getLastSequenceNumber();
572 * Persists the given Payload in the journal and replicates to any followers. After successful completion,
573 * {@link #applyState(ActorRef, Identifier, Object)} is notified.
575 * @param clientActor optional ActorRef that is provided via the applyState callback
576 * @param identifier the payload identifier
577 * @param data the payload data to persist
578 * @param batchHint if true, an attempt is made to delay immediate replication and batch the payload with
579 * subsequent payloads for efficiency. Otherwise the payload is immediately replicated.
581 protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data,
582 final boolean batchHint) {
583 ReplicatedLogEntry replicatedLogEntry = new SimpleReplicatedLogEntry(
584 context.getReplicatedLog().lastIndex() + 1,
585 context.getTermInformation().getCurrentTerm(), data);
586 replicatedLogEntry.setPersistencePending(true);
588 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
590 final RaftActorContext raftContext = getRaftActorContext();
592 boolean wasAppended = replicatedLog().appendAndPersist(replicatedLogEntry, persistedLogEntry -> {
593 // Clear the persistence pending flag in the log entry.
594 persistedLogEntry.setPersistencePending(false);
596 if (!hasFollowers()) {
597 // Increment the Commit Index and the Last Applied values
598 raftContext.setCommitIndex(persistedLogEntry.getIndex());
599 raftContext.setLastApplied(persistedLogEntry.getIndex());
601 // Apply the state immediately.
602 handleApplyState(new ApplyState(clientActor, identifier, persistedLogEntry));
604 // Send a ApplyJournalEntries message so that we write the fact that we applied
605 // the state to durable storage
606 self().tell(new ApplyJournalEntries(persistedLogEntry.getIndex()), self());
609 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
611 // Local persistence is complete so send the CheckConsensusReached message to the behavior (which
612 // normally should still be the leader) to check if consensus has now been reached in conjunction with
613 // follower replication.
614 getCurrentBehavior().handleMessage(getSelf(), CheckConsensusReached.INSTANCE);
618 if (wasAppended && hasFollowers()) {
619 // Send log entry for replication.
620 getCurrentBehavior().handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry,
625 private ReplicatedLog replicatedLog() {
626 return context.getReplicatedLog();
629 protected String getId() {
630 return context.getId();
634 void setCurrentBehavior(RaftActorBehavior behavior) {
635 context.setCurrentBehavior(behavior);
638 protected RaftActorBehavior getCurrentBehavior() {
639 return context.getCurrentBehavior();
643 * Derived actors can call the isLeader method to check if the current
644 * RaftActor is the Leader or not.
646 * @return true it this RaftActor is a Leader false otherwise
648 protected boolean isLeader() {
649 return context.getId().equals(getCurrentBehavior().getLeaderId());
652 protected final boolean isLeaderActive() {
653 return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
654 && !shuttingDown && !isLeadershipTransferInProgress();
657 private boolean isLeadershipTransferInProgress() {
658 return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
662 * Derived actor can call getLeader if they need a reference to the Leader.
663 * This would be useful for example in forwarding a request to an actor
664 * which is the leader
666 * @return A reference to the leader if known, null otherwise
668 public ActorSelection getLeader() {
669 String leaderAddress = getLeaderAddress();
671 if (leaderAddress == null) {
675 return context.actorSelection(leaderAddress);
679 * Returns the id of the current leader.
681 * @return the current leader's id
683 protected final String getLeaderId() {
684 return getCurrentBehavior().getLeaderId();
688 protected final RaftState getRaftState() {
689 return getCurrentBehavior().state();
692 protected Long getCurrentTerm() {
693 return context.getTermInformation().getCurrentTerm();
696 protected RaftActorContext getRaftActorContext() {
700 protected void updateConfigParams(ConfigParams configParams) {
702 // obtain the RaftPolicy for oldConfigParams and the updated one.
703 String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
704 String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
706 LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
707 oldRaftPolicy, newRaftPolicy);
708 context.setConfigParams(configParams);
709 if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
710 // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
711 // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
712 // avoids potential disruption. Otherwise, switch to Follower normally.
713 RaftActorBehavior behavior = getCurrentBehavior();
714 if (behavior != null && behavior.state() == RaftState.Follower) {
715 String previousLeaderId = behavior.getLeaderId();
716 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
718 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
721 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
723 initializeBehavior();
728 public final DataPersistenceProvider persistence() {
729 return delegatingPersistenceProvider.getDelegate();
732 public void setPersistence(DataPersistenceProvider provider) {
733 delegatingPersistenceProvider.setDelegate(provider);
736 protected void setPersistence(boolean persistent) {
737 DataPersistenceProvider currentPersistence = persistence();
738 if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
739 setPersistence(new PersistentDataProvider(this));
741 if (getCurrentBehavior() != null) {
742 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
745 } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
746 setPersistence(new NonPersistentDataProvider() {
748 * The way snapshotting works is,
750 * <li> RaftActor calls createSnapshot on the Shard
751 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
752 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
753 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
754 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
755 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
759 public void saveSnapshot(Object object) {
760 // Make saving Snapshot successful
761 // Committing the snapshot here would end up calling commit in the creating state which would
762 // be a state violation. That's why now we send a message to commit the snapshot.
763 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
770 * setPeerAddress sets the address of a known peer at a later time.
773 * This is to account for situations where a we know that a peer
774 * exists but we do not know an address up-front. This may also be used in
775 * situations where a known peer starts off in a different location and we
776 * need to change it's address
779 * Note that if the peerId does not match the list of peers passed to
780 * this actor during construction an IllegalStateException will be thrown.
782 protected void setPeerAddress(String peerId, String peerAddress) {
783 context.setPeerAddress(peerId, peerAddress);
787 * The applyState method will be called by the RaftActor when some data
788 * needs to be applied to the actor's state.
790 * @param clientActor A reference to the client who sent this message. This
791 * is the same reference that was passed to persistData
792 * by the derived actor. clientActor may be null when
793 * the RaftActor is behaving as a follower or during
795 * @param identifier The identifier of the persisted data. This is also
796 * the same identifier that was passed to persistData by
797 * the derived actor. identifier may be null when
798 * the RaftActor is behaving as a follower or during
800 * @param data A piece of data that was persisted by the persistData call.
801 * This should NEVER be null.
803 protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
806 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
809 protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
812 * This method is called when recovery is complete.
814 protected abstract void onRecoveryComplete();
817 * Returns the RaftActorSnapshotCohort to participate in snapshot captures.
820 protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
823 * This method will be called by the RaftActor when the state of the
824 * RaftActor changes. The derived actor can then use methods like
825 * isLeader or getLeader to do something useful
827 protected abstract void onStateChanged();
830 * Notifier Actor for this RaftActor to notify when a role change happens.
832 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
834 protected abstract Optional<ActorRef> getRoleChangeNotifier();
837 * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
838 * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
839 * work prior to performing the operation. On completion of any work, the run method must be called on the
840 * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
841 * this actor's thread dispatcher as as it modifies internal state.
844 * The default implementation immediately runs the operation.
846 * @param operation the operation to run
848 protected void pauseLeader(Runnable operation) {
852 protected void onLeaderChanged(String oldLeader, String newLeader) {
855 private String getLeaderAddress() {
857 return getSelf().path().toString();
859 String leaderId = getLeaderId();
860 if (leaderId == null) {
863 String peerAddress = context.getPeerAddress(leaderId);
864 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
869 protected boolean hasFollowers() {
870 return getRaftActorContext().hasFollowers();
873 private void captureSnapshot() {
874 SnapshotManager snapshotManager = context.getSnapshotManager();
876 if (!snapshotManager.isCapturing()) {
877 final long idx = getCurrentBehavior().getReplicatedToAllIndex();
878 LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
879 replicatedLog().last(), idx);
881 snapshotManager.capture(replicatedLog().last(), idx);
886 * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
887 * in which case we need to step down.
889 void becomeNonVoting() {
891 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
893 public void onSuccess(ActorRef raftActorRef) {
894 LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
895 ensureFollowerState();
899 public void onFailure(ActorRef raftActorRef) {
900 LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
901 ensureFollowerState();
904 private void ensureFollowerState() {
905 // Whether or not leadership transfer succeeded, we have to step down as leader and
906 // switch to Follower so ensure that.
907 if (getRaftState() != RaftState.Follower) {
908 initializeBehavior();
916 * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
918 private abstract static class BehaviorState implements Immutable {
919 @Nullable abstract RaftActorBehavior getBehavior();
921 @Nullable abstract String getLastValidLeaderId();
923 @Nullable abstract String getLastLeaderId();
925 @Nullable abstract short getLeaderPayloadVersion();
929 * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
931 private static final class SimpleBehaviorState extends BehaviorState {
932 private final RaftActorBehavior behavior;
933 private final String lastValidLeaderId;
934 private final String lastLeaderId;
935 private final short leaderPayloadVersion;
937 SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
938 final RaftActorBehavior behavior) {
939 this.lastValidLeaderId = lastValidLeaderId;
940 this.lastLeaderId = lastLeaderId;
941 this.behavior = Preconditions.checkNotNull(behavior);
942 this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
946 RaftActorBehavior getBehavior() {
951 String getLastValidLeaderId() {
952 return lastValidLeaderId;
956 short getLeaderPayloadVersion() {
957 return leaderPayloadVersion;
961 String getLastLeaderId() {
967 * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
968 * An instance is created for each RaftActor. It has two functions:
969 * - it keeps track of the last leader ID we have encountered since we have been created
970 * - it creates state capture needed to transition from one behavior to the next
972 private static final class BehaviorStateTracker {
974 * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
975 * allowed before we receive the first message, we know the leader ID to be null.
977 private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
979 RaftActorBehavior getBehavior() {
984 String getLastValidLeaderId() {
989 short getLeaderPayloadVersion() {
994 String getLastLeaderId() {
999 private String lastValidLeaderId;
1000 private String lastLeaderId;
1002 BehaviorState capture(final RaftActorBehavior behavior) {
1003 if (behavior == null) {
1004 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
1005 return NULL_BEHAVIOR_STATE;
1008 lastLeaderId = behavior.getLeaderId();
1009 if (lastLeaderId != null) {
1010 lastValidLeaderId = lastLeaderId;
1013 return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);