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
9 package org.opendaylight.controller.cluster.raft;
11 import static com.google.common.base.Verify.verify;
12 import static java.util.Objects.requireNonNull;
14 import akka.actor.ActorRef;
15 import akka.actor.ActorSelection;
16 import akka.actor.PoisonPill;
17 import akka.actor.Status;
18 import com.google.common.annotations.VisibleForTesting;
19 import java.util.ArrayList;
20 import java.util.Collection;
21 import java.util.HashMap;
22 import java.util.List;
24 import java.util.Objects;
25 import java.util.Optional;
26 import java.util.concurrent.TimeUnit;
27 import org.apache.commons.lang3.time.DurationFormatUtils;
28 import org.eclipse.jdt.annotation.NonNull;
29 import org.eclipse.jdt.annotation.Nullable;
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 boolean shuttingDown;
126 protected RaftActor(final String id, final Map<String, String> peerAddresses,
127 final Optional<ConfigParams> configParams, final short payloadVersion) {
129 persistentProvider = new PersistentDataProvider(this);
130 delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
132 context = new RaftActorContextImpl(this.getSelf(),
133 this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
134 -1, -1, peerAddresses,
135 configParams.isPresent() ? configParams.get() : new DefaultConfigParamsImpl(),
136 delegatingPersistenceProvider, this::handleApplyState, LOG, this::executeInSelf);
138 context.setPayloadVersion(payloadVersion);
139 context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
143 public void preStart() throws Exception {
144 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
145 context.getConfigParams().getJournalRecoveryLogBatchSize());
149 snapshotSupport = newRaftActorSnapshotMessageSupport();
150 serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
154 public void postStop() throws Exception {
160 protected void handleRecover(final Object message) {
161 if (raftRecovery == null) {
162 raftRecovery = newRaftActorRecoverySupport();
165 boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
166 if (recoveryComplete) {
167 onRecoveryComplete();
169 initializeBehavior();
175 protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
176 return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
180 void initializeBehavior() {
181 changeCurrentBehavior(new Follower(context));
185 @SuppressWarnings("checkstyle:IllegalCatch")
186 protected void changeCurrentBehavior(final RaftActorBehavior newBehavior) {
187 final RaftActorBehavior currentBehavior = getCurrentBehavior();
188 if (currentBehavior != null) {
190 currentBehavior.close();
191 } catch (Exception e) {
192 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
196 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
197 setCurrentBehavior(newBehavior);
198 handleBehaviorChange(state, newBehavior);
202 * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
203 * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
204 * implementation for messages which they do not handle
206 * @param message Incoming command message
208 protected void handleNonRaftCommand(final Object message) {
215 * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
216 * {@link #handleNonRaftCommand(Object)} instead.
220 // FIXME: make this method final once our unit tests do not need to override it
221 protected void handleCommand(final Object message) {
222 if (serverConfigurationSupport.handleMessage(message, getSender())) {
225 if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
228 if (message instanceof ApplyState) {
229 ApplyState applyState = (ApplyState) message;
231 if (!hasFollowers()) {
232 // for single node, the capture should happen after the apply state
233 // as we delete messages from the persistent journal which have made it to the snapshot
234 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
235 // and recovery shows data missing
236 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
238 context.getSnapshotManager().trimLog(context.getLastApplied());
241 possiblyHandleBehaviorMessage(message);
242 } else if (message instanceof ApplyJournalEntries) {
243 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
244 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
246 persistence().persistAsync(applyEntries, NoopProcedure.instance());
248 } else if (message instanceof FindLeader) {
250 new FindLeaderReply(getLeaderAddress()),
253 } else if (message instanceof GetOnDemandRaftState) {
254 onGetOnDemandRaftStats();
255 } else if (message instanceof InitiateCaptureSnapshot) {
257 } else if (message instanceof SwitchBehavior) {
258 switchBehavior((SwitchBehavior) message);
259 } else if (message instanceof LeaderTransitioning) {
260 onLeaderTransitioning((LeaderTransitioning)message);
261 } else if (message instanceof Shutdown) {
263 } else if (message instanceof Runnable) {
264 ((Runnable)message).run();
265 } else if (message instanceof NoopPayload) {
266 persistData(null, null, (NoopPayload) message, false);
267 } else if (message instanceof RequestLeadership) {
268 onRequestLeadership((RequestLeadership) message);
269 } else if (!possiblyHandleBehaviorMessage(message)) {
270 handleNonRaftCommand(message);
274 private void onRequestLeadership(final RequestLeadership message) {
275 LOG.debug("{}: onRequestLeadership {}", persistenceId(), message);
277 // non-leader cannot satisfy leadership request
278 LOG.warn("{}: onRequestLeadership {} was sent to non-leader."
279 + " Current behavior: {}. Sending failure response",
280 persistenceId(), message, getCurrentBehavior().state());
281 message.getReplyTo().tell(new LeadershipTransferFailedException("Cannot transfer leader to "
282 + message.getRequestedFollowerId()
283 + ". RequestLeadership message was sent to non-leader " + persistenceId()), getSelf());
287 final String requestedFollowerId = message.getRequestedFollowerId();
288 final ActorRef replyTo = message.getReplyTo();
289 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
291 public void onSuccess(final ActorRef raftActorRef) {
293 if (!requestedFollowerId.equals(getLeaderId())) {
294 onFailure(raftActorRef);
297 LOG.debug("{}: Leadership transferred successfully to {}", persistenceId(), requestedFollowerId);
298 replyTo.tell(new Status.Success(null), getSelf());
302 public void onFailure(final ActorRef raftActorRef) {
303 LOG.debug("{}: LeadershipTransfer request from {} failed", persistenceId(), requestedFollowerId);
304 replyTo.tell(new Status.Failure(
305 new LeadershipTransferFailedException(
306 "Failed to transfer leadership to " + requestedFollowerId
307 + ". Follower is not ready to become leader")),
310 }, message.getRequestedFollowerId(), RaftActorLeadershipTransferCohort.USE_DEFAULT_LEADER_TIMEOUT);
313 private boolean possiblyHandleBehaviorMessage(final Object message) {
314 final RaftActorBehavior currentBehavior = getCurrentBehavior();
315 final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
317 // A behavior indicates that it processed the change by returning a reference to the next behavior
318 // to be used. A null return indicates it has not processed the message and we should be passing it to
319 // the subclass for handling.
320 final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
321 if (nextBehavior != null) {
322 switchBehavior(state, nextBehavior);
329 private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete,
330 final @Nullable String followerId, final long newLeaderTimeoutInMillis) {
331 LOG.debug("{}: Initiating leader transfer", persistenceId());
333 RaftActorLeadershipTransferCohort leadershipTransferInProgress = context.getRaftActorLeadershipTransferCohort();
334 if (leadershipTransferInProgress == null) {
335 leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this, followerId);
336 leadershipTransferInProgress.setNewLeaderTimeoutInMillis(newLeaderTimeoutInMillis);
337 leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
339 public void onSuccess(final ActorRef raftActorRef) {
340 context.setRaftActorLeadershipTransferCohort(null);
344 public void onFailure(final ActorRef raftActorRef) {
345 context.setRaftActorLeadershipTransferCohort(null);
349 leadershipTransferInProgress.addOnComplete(onComplete);
351 context.setRaftActorLeadershipTransferCohort(leadershipTransferInProgress);
352 leadershipTransferInProgress.init();
355 LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
356 leadershipTransferInProgress.addOnComplete(onComplete);
360 private void onShutDown() {
361 LOG.debug("{}: onShutDown", persistenceId());
369 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
370 switch (currentBehavior.state()) {
373 // Fall-through to more work
376 // For non-leaders shutdown is a no-op
377 self().tell(PoisonPill.getInstance(), self());
381 if (context.hasFollowers()) {
382 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
384 public void onSuccess(final ActorRef raftActorRef) {
385 LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
386 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
390 public void onFailure(final ActorRef raftActorRef) {
391 LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
392 raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
394 }, null, TimeUnit.MILLISECONDS.convert(2, TimeUnit.SECONDS));
396 pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
398 protected void doRun() {
399 self().tell(PoisonPill.getInstance(), self());
403 protected void doCancel() {
404 self().tell(PoisonPill.getInstance(), self());
410 private void onLeaderTransitioning(final LeaderTransitioning leaderTransitioning) {
411 LOG.debug("{}: onLeaderTransitioning: {}", persistenceId(), leaderTransitioning);
412 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
413 if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()
414 && leaderTransitioning.getLeaderId().equals(getCurrentBehavior().getLeaderId())) {
415 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
416 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
420 private void switchBehavior(final SwitchBehavior message) {
421 if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
422 RaftState newState = message.getNewState();
423 if (newState == RaftState.Leader || newState == RaftState.Follower) {
424 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
425 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
426 AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
428 LOG.warn("Switching to behavior : {} - not supported", newState);
433 private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
434 setCurrentBehavior(nextBehavior);
435 handleBehaviorChange(oldBehaviorState, nextBehavior);
439 RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
440 return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
443 private void onGetOnDemandRaftStats() {
444 // Debugging message to retrieve raft stats.
446 Map<String, String> peerAddresses = new HashMap<>();
447 Map<String, Boolean> peerVotingStates = new HashMap<>();
448 for (PeerInfo info: context.getPeers()) {
449 peerVotingStates.put(info.getId(), info.isVoting());
450 peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
453 final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
454 OnDemandRaftState.AbstractBuilder<?, ?> builder = newOnDemandRaftStateBuilder()
455 .commitIndex(context.getCommitIndex())
456 .currentTerm(context.getTermInformation().getCurrentTerm())
457 .inMemoryJournalDataSize(replicatedLog().dataSize())
458 .inMemoryJournalLogSize(replicatedLog().size())
459 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
460 .lastApplied(context.getLastApplied())
461 .lastIndex(replicatedLog().lastIndex())
462 .lastTerm(replicatedLog().lastTerm())
463 .leader(getLeaderId())
464 .raftState(currentBehavior.state().toString())
465 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
466 .snapshotIndex(replicatedLog().getSnapshotIndex())
467 .snapshotTerm(replicatedLog().getSnapshotTerm())
468 .votedFor(context.getTermInformation().getVotedFor())
469 .isVoting(context.isVotingMember())
470 .peerAddresses(peerAddresses)
471 .peerVotingStates(peerVotingStates)
472 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
474 ReplicatedLogEntry lastLogEntry = replicatedLog().last();
475 if (lastLogEntry != null) {
476 builder.lastLogIndex(lastLogEntry.getIndex());
477 builder.lastLogTerm(lastLogEntry.getTerm());
480 if (getCurrentBehavior() instanceof AbstractLeader) {
481 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
482 Collection<String> followerIds = leader.getFollowerIds();
483 List<FollowerInfo> followerInfoList = new ArrayList<>(followerIds.size());
484 for (String id: followerIds) {
485 final FollowerLogInformation info = leader.getFollower(id);
486 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
487 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(
488 TimeUnit.NANOSECONDS.toMillis(info.nanosSinceLastActivity())),
489 context.getPeerInfo(info.getId()).isVoting()));
492 builder.followerInfoList(followerInfoList);
495 sender().tell(builder.build(), self());
499 protected OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
500 return OnDemandRaftState.builder();
503 private void handleBehaviorChange(final BehaviorState oldBehaviorState, final RaftActorBehavior currentBehavior) {
504 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
506 if (oldBehavior != currentBehavior) {
510 String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
511 String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
512 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
514 // it can happen that the state has not changed but the leader has changed.
515 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
516 if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
517 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
518 if (roleChangeNotifier.isPresent()) {
519 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
520 currentBehavior.getLeaderPayloadVersion()), getSelf());
523 onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
525 RaftActorLeadershipTransferCohort leadershipTransferInProgress =
526 context.getRaftActorLeadershipTransferCohort();
527 if (leadershipTransferInProgress != null) {
528 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
531 serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
534 if (roleChangeNotifier.isPresent()
535 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
536 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
537 currentBehavior.state().name()), getSelf());
541 private void handleApplyState(final ApplyState applyState) {
542 long startTime = System.nanoTime();
544 Payload payload = applyState.getReplicatedLogEntry().getData();
545 if (LOG.isDebugEnabled()) {
546 LOG.debug("{}: Applying state for log index {} data {}",
547 persistenceId(), applyState.getReplicatedLogEntry().getIndex(), payload);
550 if (!(payload instanceof NoopPayload) && !(payload instanceof ServerConfigurationPayload)) {
551 applyState(applyState.getClientActor(), applyState.getIdentifier(), payload);
554 long elapsedTime = System.nanoTime() - startTime;
555 if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
556 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
557 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
560 // Send the ApplyState message back to self to handle further processing asynchronously.
561 self().tell(applyState, self());
564 protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
565 final short leaderPayloadVersion) {
566 return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
570 public long snapshotSequenceNr() {
571 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
572 // so that we can delete the persistent journal based on the saved sequence-number
573 // However , when akka replays the journal during recovery, it replays it from the sequence number when the
574 // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
575 // last-sequence number known to us.
576 return context.getSnapshotManager().getLastSequenceNumber();
580 * Persists the given Payload in the journal and replicates to any followers. After successful completion,
581 * {@link #applyState(ActorRef, Identifier, Object)} is notified.
583 * @param clientActor optional ActorRef that is provided via the applyState callback
584 * @param identifier the payload identifier
585 * @param data the payload data to persist
586 * @param batchHint if true, an attempt is made to delay immediate replication and batch the payload with
587 * subsequent payloads for efficiency. Otherwise the payload is immediately replicated.
589 protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data,
590 final boolean batchHint) {
591 ReplicatedLogEntry replicatedLogEntry = new SimpleReplicatedLogEntry(
592 context.getReplicatedLog().lastIndex() + 1,
593 context.getTermInformation().getCurrentTerm(), data);
594 replicatedLogEntry.setPersistencePending(true);
596 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
598 final RaftActorContext raftContext = getRaftActorContext();
600 boolean wasAppended = replicatedLog().appendAndPersist(replicatedLogEntry, persistedLogEntry -> {
601 // Clear the persistence pending flag in the log entry.
602 persistedLogEntry.setPersistencePending(false);
604 if (!hasFollowers()) {
605 // Increment the Commit Index and the Last Applied values
606 raftContext.setCommitIndex(persistedLogEntry.getIndex());
607 raftContext.setLastApplied(persistedLogEntry.getIndex());
609 // Apply the state immediately.
610 handleApplyState(new ApplyState(clientActor, identifier, persistedLogEntry));
612 // Send a ApplyJournalEntries message so that we write the fact that we applied
613 // the state to durable storage
614 self().tell(new ApplyJournalEntries(persistedLogEntry.getIndex()), self());
617 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
619 // Local persistence is complete so send the CheckConsensusReached message to the behavior (which
620 // normally should still be the leader) to check if consensus has now been reached in conjunction with
621 // follower replication.
622 getCurrentBehavior().handleMessage(getSelf(), CheckConsensusReached.INSTANCE);
626 if (wasAppended && hasFollowers()) {
627 // Send log entry for replication.
628 getCurrentBehavior().handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry,
633 private ReplicatedLog replicatedLog() {
634 return context.getReplicatedLog();
637 protected String getId() {
638 return context.getId();
642 void setCurrentBehavior(final RaftActorBehavior behavior) {
643 context.setCurrentBehavior(behavior);
646 protected RaftActorBehavior getCurrentBehavior() {
647 return context.getCurrentBehavior();
651 * Derived actors can call the isLeader method to check if the current
652 * RaftActor is the Leader or not.
654 * @return true it this RaftActor is a Leader false otherwise
656 protected boolean isLeader() {
657 return context.getId().equals(getCurrentBehavior().getLeaderId());
660 protected final boolean isLeaderActive() {
661 return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
662 && !shuttingDown && !isLeadershipTransferInProgress();
665 protected boolean isLeadershipTransferInProgress() {
666 RaftActorLeadershipTransferCohort leadershipTransferInProgress = context.getRaftActorLeadershipTransferCohort();
667 return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
671 * Derived actor can call getLeader if they need a reference to the Leader.
672 * This would be useful for example in forwarding a request to an actor
673 * which is the leader
675 * @return A reference to the leader if known, null otherwise
677 public ActorSelection getLeader() {
678 String leaderAddress = getLeaderAddress();
680 if (leaderAddress == null) {
684 return context.actorSelection(leaderAddress);
688 * Returns the id of the current leader.
690 * @return the current leader's id
692 protected final String getLeaderId() {
693 return getCurrentBehavior().getLeaderId();
697 protected final RaftState getRaftState() {
698 return getCurrentBehavior().state();
701 protected Long getCurrentTerm() {
702 return context.getTermInformation().getCurrentTerm();
705 protected RaftActorContext getRaftActorContext() {
709 protected void updateConfigParams(final ConfigParams configParams) {
711 // obtain the RaftPolicy for oldConfigParams and the updated one.
712 String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
713 String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
715 LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
716 oldRaftPolicy, newRaftPolicy);
717 context.setConfigParams(configParams);
718 if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
719 // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
720 // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
721 // avoids potential disruption. Otherwise, switch to Follower normally.
722 RaftActorBehavior behavior = getCurrentBehavior();
723 if (behavior != null && behavior.state() == RaftState.Follower) {
724 String previousLeaderId = behavior.getLeaderId();
725 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
727 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
730 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
732 initializeBehavior();
737 public final DataPersistenceProvider persistence() {
738 return delegatingPersistenceProvider.getDelegate();
741 public void setPersistence(final DataPersistenceProvider provider) {
742 delegatingPersistenceProvider.setDelegate(provider);
745 protected void setPersistence(final boolean persistent) {
746 DataPersistenceProvider currentPersistence = persistence();
747 if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
748 setPersistence(new PersistentDataProvider(this));
750 if (getCurrentBehavior() != null) {
751 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
754 } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
755 setPersistence(new NonPersistentDataProvider(this) {
757 * The way snapshotting works is,
759 * <li> RaftActor calls createSnapshot on the Shard
760 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
761 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
762 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
763 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
764 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
768 public void saveSnapshot(final Object object) {
769 // Make saving Snapshot successful
770 // Committing the snapshot here would end up calling commit in the creating state which would
771 // be a state violation. That's why now we send a message to commit the snapshot.
772 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
779 * setPeerAddress sets the address of a known peer at a later time.
782 * This is to account for situations where a we know that a peer
783 * exists but we do not know an address up-front. This may also be used in
784 * situations where a known peer starts off in a different location and we
785 * need to change it's address
788 * Note that if the peerId does not match the list of peers passed to
789 * this actor during construction an IllegalStateException will be thrown.
791 protected void setPeerAddress(final String peerId, final String peerAddress) {
792 context.setPeerAddress(peerId, peerAddress);
796 * The applyState method will be called by the RaftActor when some data
797 * needs to be applied to the actor's state.
799 * @param clientActor A reference to the client who sent this message. This
800 * is the same reference that was passed to persistData
801 * by the derived actor. clientActor may be null when
802 * the RaftActor is behaving as a follower or during
804 * @param identifier The identifier of the persisted data. This is also
805 * the same identifier that was passed to persistData by
806 * the derived actor. identifier may be null when
807 * the RaftActor is behaving as a follower or during
809 * @param data A piece of data that was persisted by the persistData call.
810 * This should NEVER be null.
812 protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
815 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
817 protected abstract @NonNull RaftActorRecoveryCohort getRaftActorRecoveryCohort();
820 * This method is called when recovery is complete.
822 protected abstract void onRecoveryComplete();
825 * Returns the RaftActorSnapshotCohort to participate in snapshot captures.
827 protected abstract @NonNull RaftActorSnapshotCohort getRaftActorSnapshotCohort();
830 * This method will be called by the RaftActor when the state of the
831 * RaftActor changes. The derived actor can then use methods like
832 * isLeader or getLeader to do something useful
834 protected abstract void onStateChanged();
837 * Notifier Actor for this RaftActor to notify when a role change happens.
839 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
841 protected abstract Optional<ActorRef> getRoleChangeNotifier();
844 * This method is called on the leader when a voting change operation completes.
846 protected void onVotingStateChangeComplete() {
850 * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
851 * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
852 * work prior to performing the operation. On completion of any work, the run method must be called on the
853 * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
854 * this actor's thread dispatcher as as it modifies internal state.
857 * The default implementation immediately runs the operation.
859 * @param operation the operation to run
861 protected void pauseLeader(final Runnable operation) {
866 * This method is invoked when the actions hooked to the leader becoming paused failed to execute and the leader
867 * should resume normal operations.
870 * Note this method can be invoked even before the operation supplied to {@link #pauseLeader(Runnable)} is invoked.
872 protected void unpauseLeader() {
876 protected void onLeaderChanged(final String oldLeader, final String newLeader) {
879 private String getLeaderAddress() {
881 return getSelf().path().toString();
883 String leaderId = getLeaderId();
884 if (leaderId == null) {
887 String peerAddress = context.getPeerAddress(leaderId);
888 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
893 protected boolean hasFollowers() {
894 return getRaftActorContext().hasFollowers();
897 private void captureSnapshot() {
898 SnapshotManager snapshotManager = context.getSnapshotManager();
900 if (!snapshotManager.isCapturing()) {
901 final long idx = getCurrentBehavior().getReplicatedToAllIndex();
902 LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
903 replicatedLog().last(), idx);
905 snapshotManager.captureWithForcedTrim(replicatedLog().last(), idx);
910 * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
911 * in which case we need to step down.
913 void becomeNonVoting() {
915 initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
917 public void onSuccess(final ActorRef raftActorRef) {
918 LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
919 ensureFollowerState();
923 public void onFailure(final ActorRef raftActorRef) {
924 LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
925 ensureFollowerState();
928 private void ensureFollowerState() {
929 // Whether or not leadership transfer succeeded, we have to step down as leader and
930 // switch to Follower so ensure that.
931 if (getRaftState() != RaftState.Follower) {
932 initializeBehavior();
935 }, null, RaftActorLeadershipTransferCohort.USE_DEFAULT_LEADER_TIMEOUT);
940 * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
942 private abstract static class BehaviorState implements Immutable {
943 @Nullable abstract RaftActorBehavior getBehavior();
945 @Nullable abstract String getLastValidLeaderId();
947 @Nullable abstract String getLastLeaderId();
949 abstract short getLeaderPayloadVersion();
953 * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
955 private static final class SimpleBehaviorState extends BehaviorState {
956 private final RaftActorBehavior behavior;
957 private final String lastValidLeaderId;
958 private final String lastLeaderId;
959 private final short leaderPayloadVersion;
961 SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
962 final RaftActorBehavior behavior) {
963 this.lastValidLeaderId = lastValidLeaderId;
964 this.lastLeaderId = lastLeaderId;
965 this.behavior = requireNonNull(behavior);
966 this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
970 RaftActorBehavior getBehavior() {
975 String getLastValidLeaderId() {
976 return lastValidLeaderId;
980 short getLeaderPayloadVersion() {
981 return leaderPayloadVersion;
985 String getLastLeaderId() {
991 * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
992 * An instance is created for each RaftActor. It has two functions:
993 * - it keeps track of the last leader ID we have encountered since we have been created
994 * - it creates state capture needed to transition from one behavior to the next
996 private static final class BehaviorStateTracker {
998 * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
999 * allowed before we receive the first message, we know the leader ID to be null.
1001 private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
1003 RaftActorBehavior getBehavior() {
1008 String getLastValidLeaderId() {
1013 short getLeaderPayloadVersion() {
1018 String getLastLeaderId() {
1023 private String lastValidLeaderId;
1024 private String lastLeaderId;
1026 BehaviorState capture(final RaftActorBehavior behavior) {
1027 if (behavior == null) {
1028 verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
1029 return NULL_BEHAVIOR_STATE;
1032 lastLeaderId = behavior.getLeaderId();
1033 if (lastLeaderId != null) {
1034 lastValidLeaderId = lastLeaderId;
1037 return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);