opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java

   1 /*
   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.
   4  *
   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
   8  */
   9
  10 package org.opendaylight.controller.cluster.raft;
  11
  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;
  23 import java.util.Map;
  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.persisted.SimpleReplicatedLogEntry;
  55 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  56 import org.opendaylight.yangtools.concepts.Identifier;
  57 import org.opendaylight.yangtools.concepts.Immutable;
  58
  59 /**
  60  * RaftActor encapsulates a state machine that needs to be kept synchronized
  61  * in a cluster. It implements the RAFT algorithm as described in the paper
  62  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  63  * In Search of an Understandable Consensus Algorithm</a>
  64  *
  65  * <p>
  66  * RaftActor has 3 states and each state has a certain behavior associated
  67  * with it. A Raft actor can behave as,
  68  * <ul>
  69  * <li> A Leader </li>
  70  * <li> A Follower (or) </li>
  71  * <li> A Candidate </li>
  72  * </ul>
  73  *
  74  * <p>
  75  * A RaftActor MUST be a Leader in order to accept requests from clients to
  76  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  77  * a Leader it is also responsible for ensuring that all followers ultimately
  78  * have the same log and therefore the same state machine as itself.
  79  *
  80  * <p>
  81  * The current behavior of a RaftActor determines how election for leadership
  82  * is initiated and how peer RaftActors react to request for votes.
  83  *
  84  * <p>
  85  * Each RaftActor also needs to know the current election term. It uses this
  86  * information for a couple of things. One is to simply figure out who it
  87  * voted for in the last election. Another is to figure out if the message
  88  * it received to update it's state is stale.
  89  *
  90  * <p>
  91  * The RaftActor uses akka-persistence to store it's replicated log.
  92  * Furthermore through it's behaviors a Raft Actor determines
  93  * <ul>
  94  * <li> when a log entry should be persisted </li>
  95  * <li> when a log entry should be applied to the state machine (and) </li>
  96  * <li> when a snapshot should be saved </li>
  97  * </ul>
  98  */
  99 public abstract class RaftActor extends AbstractUntypedPersistentActor {
 100
 101     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
 102
 103     /**
 104      * This context should NOT be passed directly to any other actor it is
 105      * only to be consumed by the RaftActorBehaviors.
 106      */
 107     private final RaftActorContextImpl context;
 108
 109     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
 110
 111     private final PersistentDataProvider persistentProvider;
 112
 113     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
 114
 115     private RaftActorRecoverySupport raftRecovery;
 116
 117     private RaftActorSnapshotMessageSupport snapshotSupport;
 118
 119     private RaftActorServerConfigurationSupport serverConfigurationSupport;
 120
 121     private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
 122
 123     private boolean shuttingDown;
 124
 125     protected RaftActor(String id, Map<String, String> peerAddresses,
 126          Optional<ConfigParams> configParams, short payloadVersion) {
 127
 128         persistentProvider = new PersistentDataProvider(this);
 129         delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
 130
 131         context = new RaftActorContextImpl(this.getSelf(),
 132             this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
 133             -1, -1, peerAddresses,
 134             configParams.isPresent() ? configParams.get() : new DefaultConfigParamsImpl(),
 135             delegatingPersistenceProvider, LOG);
 136
 137         context.setPayloadVersion(payloadVersion);
 138         context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
 139     }
 140
 141     @Override
 142     public void preStart() throws Exception {
 143         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 144                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 145
 146         super.preStart();
 147
 148         snapshotSupport = newRaftActorSnapshotMessageSupport();
 149         serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
 150     }
 151
 152     @Override
 153     public void postStop() {
 154         context.close();
 155         super.postStop();
 156     }
 157
 158     @Override
 159     protected void handleRecover(Object message) {
 160         if (raftRecovery == null) {
 161             raftRecovery = newRaftActorRecoverySupport();
 162         }
 163
 164         boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
 165         if (recoveryComplete) {
 166             onRecoveryComplete();
 167
 168             initializeBehavior();
 169
 170             raftRecovery = null;
 171         }
 172     }
 173
 174     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
 175         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
 176     }
 177
 178     @VisibleForTesting
 179     void initializeBehavior() {
 180         changeCurrentBehavior(new Follower(context));
 181     }
 182
 183     @VisibleForTesting
 184     @SuppressWarnings("checkstyle:IllegalCatch")
 185     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
 186         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 187         if (currentBehavior != null) {
 188             try {
 189                 currentBehavior.close();
 190             } catch (Exception e) {
 191                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
 192             }
 193         }
 194
 195         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 196         setCurrentBehavior(newBehavior);
 197         handleBehaviorChange(state, newBehavior);
 198     }
 199
 200     /**
 201      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
 202      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
 203      * implementation for messages which they do not handle
 204      *
 205      * @param message Incoming command message
 206      */
 207     protected void handleNonRaftCommand(final Object message) {
 208         unhandled(message);
 209     }
 210
 211     /**
 212      * Handles a message.
 213      *
 214      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
 215      * {@link #handleNonRaftCommand(Object)} instead.
 216      */
 217     @Deprecated
 218     @Override
 219     // FIXME: make this method final once our unit tests do not need to override it
 220     protected void handleCommand(final Object message) {
 221         if (serverConfigurationSupport.handleMessage(message, getSender())) {
 222             return;
 223         }
 224         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
 225             return;
 226         }
 227
 228         if (message instanceof ApplyState) {
 229             ApplyState applyState = (ApplyState) message;
 230
 231             long startTime = System.nanoTime();
 232
 233             if (LOG.isDebugEnabled()) {
 234                 LOG.debug("{}: Applying state for log index {} data {}",
 235                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 236                     applyState.getReplicatedLogEntry().getData());
 237             }
 238
 239             if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
 240                 applyState(applyState.getClientActor(), applyState.getIdentifier(),
 241                     applyState.getReplicatedLogEntry().getData());
 242             }
 243
 244             long elapsedTime = System.nanoTime() - startTime;
 245             if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
 246                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 247                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 248             }
 249
 250             if (!hasFollowers()) {
 251                 // for single node, the capture should happen after the apply state
 252                 // as we delete messages from the persistent journal which have made it to the snapshot
 253                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
 254                 // and recovery shows data missing
 255                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
 256
 257                 context.getSnapshotManager().trimLog(context.getLastApplied());
 258             }
 259
 260             // Send it to the current behavior - some behaviors like PreLeader need to be notified of ApplyState.
 261             possiblyHandleBehaviorMessage(message);
 262
 263         } else if (message instanceof ApplyJournalEntries) {
 264             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 265             LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
 266
 267             persistence().persistAsync(applyEntries, NoopProcedure.instance());
 268
 269         } else if (message instanceof FindLeader) {
 270             getSender().tell(
 271                 new FindLeaderReply(getLeaderAddress()),
 272                 getSelf()
 273             );
 274         } else if (message instanceof GetOnDemandRaftState) {
 275             onGetOnDemandRaftStats();
 276         } else if (message instanceof InitiateCaptureSnapshot) {
 277             captureSnapshot();
 278         } else if (message instanceof SwitchBehavior) {
 279             switchBehavior((SwitchBehavior) message);
 280         } else if (message instanceof LeaderTransitioning) {
 281             onLeaderTransitioning();
 282         } else if (message instanceof Shutdown) {
 283             onShutDown();
 284         } else if (message instanceof Runnable) {
 285             ((Runnable)message).run();
 286         } else if (message instanceof NoopPayload) {
 287             persistData(null, null, (NoopPayload)message, false);
 288         } else if (!possiblyHandleBehaviorMessage(message)) {
 289             handleNonRaftCommand(message);
 290         }
 291     }
 292
 293     private boolean possiblyHandleBehaviorMessage(final Object message) {
 294         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 295         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 296
 297         // A behavior indicates that it processed the change by returning a reference to the next behavior
 298         // to be used. A null return indicates it has not processed the message and we should be passing it to
 299         // the subclass for handling.
 300         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 301         if (nextBehavior != null) {
 302             switchBehavior(state, nextBehavior);
 303             return true;
 304         }
 305
 306         return false;
 307     }
 308
 309     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
 310         LOG.debug("{}: Initiating leader transfer", persistenceId());
 311
 312         if (leadershipTransferInProgress == null) {
 313             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
 314             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 315                 @Override
 316                 public void onSuccess(ActorRef raftActorRef) {
 317                     leadershipTransferInProgress = null;
 318                 }
 319
 320                 @Override
 321                 public void onFailure(ActorRef raftActorRef) {
 322                     leadershipTransferInProgress = null;
 323                 }
 324             });
 325
 326             leadershipTransferInProgress.addOnComplete(onComplete);
 327             leadershipTransferInProgress.init();
 328         } else {
 329             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 330             leadershipTransferInProgress.addOnComplete(onComplete);
 331         }
 332     }
 333
 334     private void onShutDown() {
 335         LOG.debug("{}: onShutDown", persistenceId());
 336
 337         if (shuttingDown) {
 338             return;
 339         }
 340
 341         shuttingDown = true;
 342
 343         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 344         if (currentBehavior.state() != RaftState.Leader) {
 345             // For non-leaders shutdown is a no-op
 346             self().tell(PoisonPill.getInstance(), self());
 347             return;
 348         }
 349
 350         if (context.hasFollowers()) {
 351             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 352                 @Override
 353                 public void onSuccess(ActorRef raftActorRef) {
 354                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 355                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 356                 }
 357
 358                 @Override
 359                 public void onFailure(ActorRef raftActorRef) {
 360                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 361                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 362                 }
 363             });
 364         } else {
 365             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 366                 @Override
 367                 protected void doRun() {
 368                     self().tell(PoisonPill.getInstance(), self());
 369                 }
 370
 371                 @Override
 372                 protected void doCancel() {
 373                     self().tell(PoisonPill.getInstance(), self());
 374                 }
 375             });
 376         }
 377     }
 378
 379     private void onLeaderTransitioning() {
 380         LOG.debug("{}: onLeaderTransitioning", persistenceId());
 381         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 382         if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
 383             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 384                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 385         }
 386     }
 387
 388     private void switchBehavior(SwitchBehavior message) {
 389         if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 390             RaftState newState = message.getNewState();
 391             if ( newState == RaftState.Leader || newState == RaftState.Follower) {
 392                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 393                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 394                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 395             } else {
 396                 LOG.warn("Switching to behavior : {} - not supported", newState);
 397             }
 398         }
 399     }
 400
 401     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 402         setCurrentBehavior(nextBehavior);
 403         handleBehaviorChange(oldBehaviorState, nextBehavior);
 404     }
 405
 406     @VisibleForTesting
 407     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 408         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 409     }
 410
 411     private void onGetOnDemandRaftStats() {
 412         // Debugging message to retrieve raft stats.
 413
 414         Map<String, String> peerAddresses = new HashMap<>();
 415         Map<String, Boolean> peerVotingStates = new HashMap<>();
 416         for (PeerInfo info: context.getPeers()) {
 417             peerVotingStates.put(info.getId(), info.isVoting());
 418             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
 419         }
 420
 421         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 422         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 423                 .commitIndex(context.getCommitIndex())
 424                 .currentTerm(context.getTermInformation().getCurrentTerm())
 425                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 426                 .inMemoryJournalLogSize(replicatedLog().size())
 427                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 428                 .lastApplied(context.getLastApplied())
 429                 .lastIndex(replicatedLog().lastIndex())
 430                 .lastTerm(replicatedLog().lastTerm())
 431                 .leader(getLeaderId())
 432                 .raftState(currentBehavior.state().toString())
 433                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 434                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 435                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 436                 .votedFor(context.getTermInformation().getVotedFor())
 437                 .isVoting(context.isVotingMember())
 438                 .peerAddresses(peerAddresses)
 439                 .peerVotingStates(peerVotingStates)
 440                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 441
 442         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 443         if (lastLogEntry != null) {
 444             builder.lastLogIndex(lastLogEntry.getIndex());
 445             builder.lastLogTerm(lastLogEntry.getTerm());
 446         }
 447
 448         if (getCurrentBehavior() instanceof AbstractLeader) {
 449             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 450             Collection<String> followerIds = leader.getFollowerIds();
 451             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 452             for (String id: followerIds) {
 453                 final FollowerLogInformation info = leader.getFollower(id);
 454                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 455                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
 456                         context.getPeerInfo(info.getId()).isVoting()));
 457             }
 458
 459             builder.followerInfoList(followerInfoList);
 460         }
 461
 462         sender().tell(builder.build(), self());
 463
 464     }
 465
 466     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
 467         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 468
 469         if (oldBehavior != currentBehavior) {
 470             onStateChanged();
 471         }
 472
 473         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
 474         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 475         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 476
 477         // it can happen that the state has not changed but the leader has changed.
 478         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 479         if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
 480                 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 481             if (roleChangeNotifier.isPresent()) {
 482                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 483                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 484             }
 485
 486             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 487
 488             if (leadershipTransferInProgress != null) {
 489                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 490             }
 491
 492             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 493         }
 494
 495         if (roleChangeNotifier.isPresent()
 496                 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
 497             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 498                     currentBehavior.state().name()), getSelf());
 499         }
 500     }
 501
 502     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 503         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 504     }
 505
 506     @Override
 507     public long snapshotSequenceNr() {
 508         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 509         // so that we can delete the persistent journal based on the saved sequence-number
 510         // However , when akka replays the journal during recovery, it replays it from the sequence number when the
 511         // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
 512         // last-sequence number known to us.
 513         return context.getSnapshotManager().getLastSequenceNumber();
 514     }
 515
 516     /**
 517      * Persists the given Payload in the journal and replicates to any followers. After successful completion,
 518      * {@link #applyState(ActorRef, Identifier, Object)} is notified.
 519      *
 520      * @param clientActor optional ActorRef that is provided via the applyState callback
 521      * @param identifier the payload identifier
 522      * @param data the payload data to persist
 523      * @param batchHint if true, an attempt is made to delay immediate replication and batch the payload with
 524      *        subsequent payloads for efficiency. Otherwise the payload is immediately replicated.
 525      */
 526     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data,
 527             final boolean batchHint) {
 528         ReplicatedLogEntry replicatedLogEntry = new SimpleReplicatedLogEntry(
 529             context.getReplicatedLog().lastIndex() + 1,
 530             context.getTermInformation().getCurrentTerm(), data);
 531         replicatedLogEntry.setPersistencePending(true);
 532
 533         LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 534
 535         final RaftActorContext raftContext = getRaftActorContext();
 536
 537         boolean wasAppended = replicatedLog().appendAndPersist(replicatedLogEntry, persistedLogEntry -> {
 538             // Clear the persistence pending flag in the log entry.
 539             persistedLogEntry.setPersistencePending(false);
 540
 541             if (!hasFollowers()) {
 542                 // Increment the Commit Index and the Last Applied values
 543                 raftContext.setCommitIndex(persistedLogEntry.getIndex());
 544                 raftContext.setLastApplied(persistedLogEntry.getIndex());
 545
 546                 // Apply the state immediately.
 547                 self().tell(new ApplyState(clientActor, identifier, persistedLogEntry), self());
 548
 549                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 550                 // the state to durable storage
 551                 self().tell(new ApplyJournalEntries(persistedLogEntry.getIndex()), self());
 552
 553             } else {
 554                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
 555
 556                 // Local persistence is complete so send the CheckConsensusReached message to the behavior (which
 557                 // normally should still be the leader) to check if consensus has now been reached in conjunction with
 558                 // follower replication.
 559                 getCurrentBehavior().handleMessage(getSelf(), CheckConsensusReached.INSTANCE);
 560             }
 561         }, true);
 562
 563         if (wasAppended && hasFollowers()) {
 564             // Send log entry for replication.
 565             getCurrentBehavior().handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry,
 566                     !batchHint));
 567         }
 568     }
 569
 570     private ReplicatedLog replicatedLog() {
 571         return context.getReplicatedLog();
 572     }
 573
 574     protected String getId() {
 575         return context.getId();
 576     }
 577
 578     @VisibleForTesting
 579     void setCurrentBehavior(RaftActorBehavior behavior) {
 580         context.setCurrentBehavior(behavior);
 581     }
 582
 583     protected RaftActorBehavior getCurrentBehavior() {
 584         return context.getCurrentBehavior();
 585     }
 586
 587     /**
 588      * Derived actors can call the isLeader method to check if the current
 589      * RaftActor is the Leader or not.
 590      *
 591      * @return true it this RaftActor is a Leader false otherwise
 592      */
 593     protected boolean isLeader() {
 594         return context.getId().equals(getCurrentBehavior().getLeaderId());
 595     }
 596
 597     protected final boolean isLeaderActive() {
 598         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
 599                 && !shuttingDown && !isLeadershipTransferInProgress();
 600     }
 601
 602     private boolean isLeadershipTransferInProgress() {
 603         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 604     }
 605
 606     /**
 607      * Derived actor can call getLeader if they need a reference to the Leader.
 608      * This would be useful for example in forwarding a request to an actor
 609      * which is the leader
 610      *
 611      * @return A reference to the leader if known, null otherwise
 612      */
 613     public ActorSelection getLeader() {
 614         String leaderAddress = getLeaderAddress();
 615
 616         if (leaderAddress == null) {
 617             return null;
 618         }
 619
 620         return context.actorSelection(leaderAddress);
 621     }
 622
 623     /**
 624      * Returns the id of the current leader.
 625      *
 626      * @return the current leader's id
 627      */
 628     protected final String getLeaderId() {
 629         return getCurrentBehavior().getLeaderId();
 630     }
 631
 632     @VisibleForTesting
 633     protected final RaftState getRaftState() {
 634         return getCurrentBehavior().state();
 635     }
 636
 637     protected Long getCurrentTerm() {
 638         return context.getTermInformation().getCurrentTerm();
 639     }
 640
 641     protected RaftActorContext getRaftActorContext() {
 642         return context;
 643     }
 644
 645     protected void updateConfigParams(ConfigParams configParams) {
 646
 647         // obtain the RaftPolicy for oldConfigParams and the updated one.
 648         String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
 649         String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
 650
 651         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 652             oldRaftPolicy, newRaftPolicy);
 653         context.setConfigParams(configParams);
 654         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 655             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 656             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 657             // avoids potential disruption. Otherwise, switch to Follower normally.
 658             RaftActorBehavior behavior = getCurrentBehavior();
 659             if (behavior != null && behavior.state() == RaftState.Follower) {
 660                 String previousLeaderId = behavior.getLeaderId();
 661                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 662
 663                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
 664                         previousLeaderId);
 665
 666                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 667             } else {
 668                 initializeBehavior();
 669             }
 670         }
 671     }
 672
 673     public final DataPersistenceProvider persistence() {
 674         return delegatingPersistenceProvider.getDelegate();
 675     }
 676
 677     public void setPersistence(DataPersistenceProvider provider) {
 678         delegatingPersistenceProvider.setDelegate(provider);
 679     }
 680
 681     protected void setPersistence(boolean persistent) {
 682         DataPersistenceProvider currentPersistence = persistence();
 683         if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
 684             setPersistence(new PersistentDataProvider(this));
 685
 686             if (getCurrentBehavior() != null) {
 687                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
 688                 captureSnapshot();
 689             }
 690         } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
 691             setPersistence(new NonPersistentDataProvider() {
 692                 /**
 693                  * The way snapshotting works is,
 694                  * <ol>
 695                  * <li> RaftActor calls createSnapshot on the Shard
 696                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 697                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 698                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 699                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 700                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 701                  * </ol>
 702                  */
 703                 @Override
 704                 public void saveSnapshot(Object object) {
 705                     // Make saving Snapshot successful
 706                     // Committing the snapshot here would end up calling commit in the creating state which would
 707                     // be a state violation. That's why now we send a message to commit the snapshot.
 708                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 709                 }
 710             });
 711         }
 712     }
 713
 714     /**
 715      * setPeerAddress sets the address of a known peer at a later time.
 716      *
 717      * <p>
 718      * This is to account for situations where a we know that a peer
 719      * exists but we do not know an address up-front. This may also be used in
 720      * situations where a known peer starts off in a different location and we
 721      * need to change it's address
 722      *
 723      * <p>
 724      * Note that if the peerId does not match the list of peers passed to
 725      * this actor during construction an IllegalStateException will be thrown.
 726      */
 727     protected void setPeerAddress(String peerId, String peerAddress) {
 728         context.setPeerAddress(peerId, peerAddress);
 729     }
 730
 731     /**
 732      * The applyState method will be called by the RaftActor when some data
 733      * needs to be applied to the actor's state.
 734      *
 735      * @param clientActor A reference to the client who sent this message. This
 736      *                    is the same reference that was passed to persistData
 737      *                    by the derived actor. clientActor may be null when
 738      *                    the RaftActor is behaving as a follower or during
 739      *                    recovery.
 740      * @param identifier  The identifier of the persisted data. This is also
 741      *                    the same identifier that was passed to persistData by
 742      *                    the derived actor. identifier may be null when
 743      *                    the RaftActor is behaving as a follower or during
 744      *                    recovery
 745      * @param data        A piece of data that was persisted by the persistData call.
 746      *                    This should NEVER be null.
 747      */
 748     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 749
 750     /**
 751      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 752      */
 753     @Nonnull
 754     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 755
 756     /**
 757      * This method is called when recovery is complete.
 758      */
 759     protected abstract void onRecoveryComplete();
 760
 761     /**
 762      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
 763      */
 764     @Nonnull
 765     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 766
 767     /**
 768      * This method will be called by the RaftActor when the state of the
 769      * RaftActor changes. The derived actor can then use methods like
 770      * isLeader or getLeader to do something useful
 771      */
 772     protected abstract void onStateChanged();
 773
 774     /**
 775      * Notifier Actor for this RaftActor to notify when a role change happens.
 776      *
 777      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 778      */
 779     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 780
 781     /**
 782      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 783      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 784      * work prior to performing the operation. On completion of any work, the run method must be called on the
 785      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 786      * this actor's thread dispatcher as as it modifies internal state.
 787      *
 788      * <p>
 789      * The default implementation immediately runs the operation.
 790      *
 791      * @param operation the operation to run
 792      */
 793     protected void pauseLeader(Runnable operation) {
 794         operation.run();
 795     }
 796
 797     protected void onLeaderChanged(String oldLeader, String newLeader) {
 798     }
 799
 800     private String getLeaderAddress() {
 801         if (isLeader()) {
 802             return getSelf().path().toString();
 803         }
 804         String leaderId = getLeaderId();
 805         if (leaderId == null) {
 806             return null;
 807         }
 808         String peerAddress = context.getPeerAddress(leaderId);
 809         LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
 810
 811         return peerAddress;
 812     }
 813
 814     protected boolean hasFollowers() {
 815         return getRaftActorContext().hasFollowers();
 816     }
 817
 818     private void captureSnapshot() {
 819         SnapshotManager snapshotManager = context.getSnapshotManager();
 820
 821         if (!snapshotManager.isCapturing()) {
 822             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 823             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 824                 replicatedLog().last(), idx);
 825
 826             snapshotManager.capture(replicatedLog().last(), idx);
 827         }
 828     }
 829
 830     /**
 831      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 832      * in which case we need to step down.
 833      */
 834     void becomeNonVoting() {
 835         if (isLeader()) {
 836             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 837                 @Override
 838                 public void onSuccess(ActorRef raftActorRef) {
 839                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 840                     ensureFollowerState();
 841                 }
 842
 843                 @Override
 844                 public void onFailure(ActorRef raftActorRef) {
 845                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 846                     ensureFollowerState();
 847                 }
 848
 849                 private void ensureFollowerState() {
 850                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 851                     // switch to Follower so ensure that.
 852                     if (getRaftState() != RaftState.Follower) {
 853                         initializeBehavior();
 854                     }
 855                 }
 856             });
 857         }
 858     }
 859
 860     /**
 861      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 862      */
 863     private abstract static class BehaviorState implements Immutable {
 864         @Nullable abstract RaftActorBehavior getBehavior();
 865
 866         @Nullable abstract String getLastValidLeaderId();
 867
 868         @Nullable abstract String getLastLeaderId();
 869
 870         @Nullable abstract short getLeaderPayloadVersion();
 871     }
 872
 873     /**
 874      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
 875      */
 876     private static final class SimpleBehaviorState extends BehaviorState {
 877         private final RaftActorBehavior behavior;
 878         private final String lastValidLeaderId;
 879         private final String lastLeaderId;
 880         private final short leaderPayloadVersion;
 881
 882         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
 883                 final RaftActorBehavior behavior) {
 884             this.lastValidLeaderId = lastValidLeaderId;
 885             this.lastLeaderId = lastLeaderId;
 886             this.behavior = Preconditions.checkNotNull(behavior);
 887             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 888         }
 889
 890         @Override
 891         RaftActorBehavior getBehavior() {
 892             return behavior;
 893         }
 894
 895         @Override
 896         String getLastValidLeaderId() {
 897             return lastValidLeaderId;
 898         }
 899
 900         @Override
 901         short getLeaderPayloadVersion() {
 902             return leaderPayloadVersion;
 903         }
 904
 905         @Override
 906         String getLastLeaderId() {
 907             return lastLeaderId;
 908         }
 909     }
 910
 911     /**
 912      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 913      * An instance is created for each RaftActor. It has two functions:
 914      * - it keeps track of the last leader ID we have encountered since we have been created
 915      * - it creates state capture needed to transition from one behavior to the next
 916      */
 917     private static final class BehaviorStateTracker {
 918         /**
 919          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
 920          * allowed before we receive the first message, we know the leader ID to be null.
 921          */
 922         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
 923             @Override
 924             RaftActorBehavior getBehavior() {
 925                 return null;
 926             }
 927
 928             @Override
 929             String getLastValidLeaderId() {
 930                 return null;
 931             }
 932
 933             @Override
 934             short getLeaderPayloadVersion() {
 935                 return -1;
 936             }
 937
 938             @Override
 939             String getLastLeaderId() {
 940                 return null;
 941             }
 942         };
 943
 944         private String lastValidLeaderId;
 945         private String lastLeaderId;
 946
 947         BehaviorState capture(final RaftActorBehavior behavior) {
 948             if (behavior == null) {
 949                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
 950                 return NULL_BEHAVIOR_STATE;
 951             }
 952
 953             lastLeaderId = behavior.getLeaderId();
 954             if (lastLeaderId != null) {
 955                 lastValidLeaderId = lastLeaderId;
 956             }
 957
 958             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
 959         }
 960     }
 961
 962 }