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.InitiateCaptureSnapshot;
  38 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
  39 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  40 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
  41 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
  42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  43 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  44 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  45 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  47 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
  48 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
  49 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
  50 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
  51 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
  52 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
  53 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  54 import org.opendaylight.yangtools.concepts.Identifier;
  55 import org.opendaylight.yangtools.concepts.Immutable;
  56 import org.slf4j.Logger;
  57 import org.slf4j.LoggerFactory;
  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  * <p/>
  65  * RaftActor has 3 states and each state has a certain behavior associated
  66  * with it. A Raft actor can behave as,
  67  * <ul>
  68  * <li> A Leader </li>
  69  * <li> A Follower (or) </li>
  70  * <li> A Candidate </li>
  71  * </ul>
  72  * <p/>
  73  * <p/>
  74  * A RaftActor MUST be a Leader in order to accept requests from clients to
  75  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  76  * a Leader it is also responsible for ensuring that all followers ultimately
  77  * have the same log and therefore the same state machine as itself.
  78  * <p/>
  79  * <p/>
  80  * The current behavior of a RaftActor determines how election for leadership
  81  * is initiated and how peer RaftActors react to request for votes.
  82  * <p/>
  83  * <p/>
  84  * Each RaftActor also needs to know the current election term. It uses this
  85  * information for a couple of things. One is to simply figure out who it
  86  * voted for in the last election. Another is to figure out if the message
  87  * it received to update it's state is stale.
  88  * <p/>
  89  * <p/>
  90  * The RaftActor uses akka-persistence to store it's replicated log.
  91  * Furthermore through it's behaviors a Raft Actor determines
  92  * <p/>
  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     protected final Logger LOG = LoggerFactory.getLogger(getClass());
 104
 105     /**
 106      * This context should NOT be passed directly to any other actor it is
 107      * only to be consumed by the RaftActorBehaviors
 108      */
 109     private final RaftActorContextImpl context;
 110
 111     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
 112
 113     private final PersistentDataProvider persistentProvider;
 114
 115     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
 116
 117     private RaftActorRecoverySupport raftRecovery;
 118
 119     private RaftActorSnapshotMessageSupport snapshotSupport;
 120
 121     private RaftActorServerConfigurationSupport serverConfigurationSupport;
 122
 123     private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
 124
 125     private boolean shuttingDown;
 126
 127     public RaftActor(String id, Map<String, String> peerAddresses,
 128          Optional<ConfigParams> configParams, short payloadVersion) {
 129
 130         persistentProvider = new PersistentDataProvider(this);
 131         delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
 132
 133         context = new RaftActorContextImpl(this.getSelf(),
 134             this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
 135             -1, -1, peerAddresses,
 136             configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl(),
 137             delegatingPersistenceProvider, LOG);
 138
 139         context.setPayloadVersion(payloadVersion);
 140         context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
 141     }
 142
 143     @Override
 144     public void preStart() throws Exception {
 145         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 146                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 147
 148         super.preStart();
 149
 150         snapshotSupport = newRaftActorSnapshotMessageSupport();
 151         serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
 152     }
 153
 154     @Override
 155     public void postStop() {
 156         context.close();
 157         super.postStop();
 158     }
 159
 160     @Override
 161     protected void handleRecover(Object message) {
 162         if(raftRecovery == null) {
 163             raftRecovery = newRaftActorRecoverySupport();
 164         }
 165
 166         boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
 167         if(recoveryComplete) {
 168             onRecoveryComplete();
 169
 170             initializeBehavior();
 171
 172             raftRecovery = null;
 173         }
 174     }
 175
 176     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
 177         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
 178     }
 179
 180     @VisibleForTesting
 181     void initializeBehavior(){
 182         changeCurrentBehavior(new Follower(context));
 183     }
 184
 185     @VisibleForTesting
 186     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
 187         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 188         if (currentBehavior != null) {
 189             try {
 190                 currentBehavior.close();
 191             } catch (Exception e) {
 192                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
 193             }
 194         }
 195
 196         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 197         setCurrentBehavior(newBehavior);
 198         handleBehaviorChange(state, newBehavior);
 199     }
 200
 201     /**
 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
 205      *
 206      * @param message Incoming command message
 207      */
 208     protected void handleNonRaftCommand(final Object message) {
 209         unhandled(message);
 210     }
 211
 212     /**
 213      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
 214      * {@link #handleNonRaftCommand(Object)} instead.
 215      */
 216     @Deprecated
 217     @Override
 218     // FIXME: make this method final once our unit tests do not need to override it
 219     protected void handleCommand(final Object message) {
 220         if (serverConfigurationSupport.handleMessage(message, getSender())) {
 221             return;
 222         }
 223         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
 224             return;
 225         }
 226
 227         if (message instanceof ApplyState) {
 228             ApplyState applyState = (ApplyState) message;
 229
 230             long startTime = System.nanoTime();
 231
 232             if(LOG.isDebugEnabled()) {
 233                 LOG.debug("{}: Applying state for log index {} data {}",
 234                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 235                     applyState.getReplicatedLogEntry().getData());
 236             }
 237
 238             if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
 239                 applyState(applyState.getClientActor(), applyState.getIdentifier(),
 240                     applyState.getReplicatedLogEntry().getData());
 241             }
 242
 243             long elapsedTime = System.nanoTime() - startTime;
 244             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 245                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 246                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 247             }
 248
 249             if (!hasFollowers()) {
 250                 // for single node, the capture should happen after the apply state
 251                 // as we delete messages from the persistent journal which have made it to the snapshot
 252                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
 253                 // and recovery shows data missing
 254                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
 255
 256                 context.getSnapshotManager().trimLog(context.getLastApplied());
 257             }
 258
 259             // Send it to the current behavior - some behaviors like PreLeader need to be notified of ApplyState.
 260             possiblyHandleBehaviorMessage(message);
 261
 262         } else if (message instanceof ApplyJournalEntries) {
 263             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 264             if(LOG.isDebugEnabled()) {
 265                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
 266             }
 267
 268             persistence().persist(applyEntries, NoopProcedure.instance());
 269
 270         } else if (message instanceof FindLeader) {
 271             getSender().tell(
 272                 new FindLeaderReply(getLeaderAddress()),
 273                 getSelf()
 274             );
 275         } else if(message instanceof GetOnDemandRaftState) {
 276             onGetOnDemandRaftStats();
 277         } else if(message instanceof InitiateCaptureSnapshot) {
 278             captureSnapshot();
 279         } else if(message instanceof SwitchBehavior) {
 280             switchBehavior((SwitchBehavior) message);
 281         } else if(message instanceof LeaderTransitioning) {
 282             onLeaderTransitioning();
 283         } else if(message instanceof Shutdown) {
 284             onShutDown();
 285         } else if(message instanceof Runnable) {
 286             ((Runnable)message).run();
 287         } else if(message instanceof NoopPayload) {
 288             persistData(null, null, (NoopPayload)message);
 289         } else if (!possiblyHandleBehaviorMessage(message)) {
 290             handleNonRaftCommand(message);
 291         }
 292     }
 293
 294     private boolean possiblyHandleBehaviorMessage(final Object message) {
 295         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 296         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 297
 298         // A behavior indicates that it processed the change by returning a reference to the next behavior
 299         // to be used. A null return indicates it has not processed the message and we should be passing it to
 300         // the subclass for handling.
 301         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 302         if (nextBehavior != null) {
 303             switchBehavior(state, nextBehavior);
 304             return true;
 305         }
 306
 307         return false;
 308     }
 309
 310     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
 311         LOG.debug("{}: Initiating leader transfer", persistenceId());
 312
 313         if(leadershipTransferInProgress == null) {
 314             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
 315             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 316                 @Override
 317                 public void onSuccess(ActorRef raftActorRef) {
 318                     leadershipTransferInProgress = null;
 319                 }
 320
 321                 @Override
 322                 public void onFailure(ActorRef raftActorRef) {
 323                     leadershipTransferInProgress = null;
 324                 }
 325             });
 326
 327             leadershipTransferInProgress.addOnComplete(onComplete);
 328             leadershipTransferInProgress.init();
 329         } else {
 330             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 331             leadershipTransferInProgress.addOnComplete(onComplete);
 332         }
 333     }
 334
 335     private void onShutDown() {
 336         LOG.debug("{}: onShutDown", persistenceId());
 337
 338         if(shuttingDown) {
 339             return;
 340         }
 341
 342         shuttingDown = true;
 343
 344         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 345         if (currentBehavior.state() != RaftState.Leader) {
 346             // For non-leaders shutdown is a no-op
 347             self().tell(PoisonPill.getInstance(), self());
 348             return;
 349         }
 350
 351         if (context.hasFollowers()) {
 352             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 353                 @Override
 354                 public void onSuccess(ActorRef raftActorRef) {
 355                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 356                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 357                 }
 358
 359                 @Override
 360                 public void onFailure(ActorRef raftActorRef) {
 361                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 362                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 363                 }
 364             });
 365         } else {
 366             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 367                 @Override
 368                 protected void doRun() {
 369                     self().tell(PoisonPill.getInstance(), self());
 370                 }
 371
 372                 @Override
 373                 protected void doCancel() {
 374                     self().tell(PoisonPill.getInstance(), self());
 375                 }
 376             });
 377         }
 378     }
 379
 380     private void onLeaderTransitioning() {
 381         LOG.debug("{}: onLeaderTransitioning", persistenceId());
 382         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 383         if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
 384             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 385                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 386         }
 387     }
 388
 389     private void switchBehavior(SwitchBehavior message) {
 390         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 391             RaftState newState = message.getNewState();
 392             if( newState == RaftState.Leader || newState == RaftState.Follower) {
 393                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 394                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 395                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 396             } else {
 397                 LOG.warn("Switching to behavior : {} - not supported", newState);
 398             }
 399         }
 400     }
 401
 402     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 403         setCurrentBehavior(nextBehavior);
 404         handleBehaviorChange(oldBehaviorState, nextBehavior);
 405     }
 406
 407     @VisibleForTesting
 408     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 409         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 410     }
 411
 412     private void onGetOnDemandRaftStats() {
 413         // Debugging message to retrieve raft stats.
 414
 415         Map<String, String> peerAddresses = new HashMap<>();
 416         Map<String, Boolean> peerVotingStates = new HashMap<>();
 417         for(PeerInfo info: context.getPeers()) {
 418             peerVotingStates.put(info.getId(), info.isVoting());
 419             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
 420         }
 421
 422         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 423         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 424                 .commitIndex(context.getCommitIndex())
 425                 .currentTerm(context.getTermInformation().getCurrentTerm())
 426                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 427                 .inMemoryJournalLogSize(replicatedLog().size())
 428                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 429                 .lastApplied(context.getLastApplied())
 430                 .lastIndex(replicatedLog().lastIndex())
 431                 .lastTerm(replicatedLog().lastTerm())
 432                 .leader(getLeaderId())
 433                 .raftState(currentBehavior.state().toString())
 434                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 435                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 436                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 437                 .votedFor(context.getTermInformation().getVotedFor())
 438                 .isVoting(context.isVotingMember())
 439                 .peerAddresses(peerAddresses)
 440                 .peerVotingStates(peerVotingStates)
 441                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 442
 443         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 444         if (lastLogEntry != null) {
 445             builder.lastLogIndex(lastLogEntry.getIndex());
 446             builder.lastLogTerm(lastLogEntry.getTerm());
 447         }
 448
 449         if(getCurrentBehavior() instanceof AbstractLeader) {
 450             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 451             Collection<String> followerIds = leader.getFollowerIds();
 452             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 453             for(String id: followerIds) {
 454                 final FollowerLogInformation info = leader.getFollower(id);
 455                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 456                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
 457                         context.getPeerInfo(info.getId()).isVoting()));
 458             }
 459
 460             builder.followerInfoList(followerInfoList);
 461         }
 462
 463         sender().tell(builder.build(), self());
 464
 465     }
 466
 467     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
 468         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 469
 470         if (oldBehavior != currentBehavior){
 471             onStateChanged();
 472         }
 473
 474         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
 475         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 476         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 477
 478         // it can happen that the state has not changed but the leader has changed.
 479         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 480         if(!Objects.equals(lastLeaderId, currentBehavior.getLeaderId()) ||
 481            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 482             if(roleChangeNotifier.isPresent()) {
 483                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 484                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 485             }
 486
 487             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 488
 489             if(leadershipTransferInProgress != null) {
 490                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 491             }
 492
 493             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 494         }
 495
 496         if (roleChangeNotifier.isPresent() &&
 497                 (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
 498             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 499                     currentBehavior.state().name()), getSelf());
 500         }
 501     }
 502
 503     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 504         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 505     }
 506
 507     @Override
 508     public long snapshotSequenceNr() {
 509         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 510         // so that we can delete the persistent journal based on the saved sequence-number
 511         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
 512         // was saved and not the number we saved.
 513         // We would want to override it , by asking akka to use the last-sequence number known to us.
 514         return context.getSnapshotManager().getLastSequenceNumber();
 515     }
 516
 517     /**
 518      * When a derived RaftActor needs to persist something it must call
 519      * persistData.
 520      *
 521      * @param clientActor
 522      * @param identifier
 523      * @param data
 524      */
 525     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
 526
 527         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 528             context.getReplicatedLog().lastIndex() + 1,
 529             context.getTermInformation().getCurrentTerm(), data);
 530
 531         if(LOG.isDebugEnabled()) {
 532             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 533         }
 534
 535         final RaftActorContext raftContext = getRaftActorContext();
 536
 537         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
 538             if (!hasFollowers()){
 539                 // Increment the Commit Index and the Last Applied values
 540                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
 541                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
 542
 543                 // Apply the state immediately.
 544                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
 545
 546                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 547                 // the state to durable storage
 548                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
 549
 550             } else {
 551                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
 552
 553                 // Send message for replication
 554                 getCurrentBehavior().handleMessage(getSelf(),
 555                         new Replicate(clientActor, identifier, replicatedLogEntry1));
 556             }
 557         });
 558     }
 559
 560     private ReplicatedLog replicatedLog() {
 561         return context.getReplicatedLog();
 562     }
 563
 564     protected String getId() {
 565         return context.getId();
 566     }
 567
 568     @VisibleForTesting
 569     void setCurrentBehavior(RaftActorBehavior behavior) {
 570         context.setCurrentBehavior(behavior);
 571     }
 572
 573     protected RaftActorBehavior getCurrentBehavior() {
 574         return context.getCurrentBehavior();
 575     }
 576
 577     /**
 578      * Derived actors can call the isLeader method to check if the current
 579      * RaftActor is the Leader or not
 580      *
 581      * @return true it this RaftActor is a Leader false otherwise
 582      */
 583     protected boolean isLeader() {
 584         return context.getId().equals(getCurrentBehavior().getLeaderId());
 585     }
 586
 587     protected final boolean isLeaderActive() {
 588         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader &&
 589                 !shuttingDown && !isLeadershipTransferInProgress();
 590     }
 591
 592     private boolean isLeadershipTransferInProgress() {
 593         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 594     }
 595
 596     /**
 597      * Derived actor can call getLeader if they need a reference to the Leader.
 598      * This would be useful for example in forwarding a request to an actor
 599      * which is the leader
 600      *
 601      * @return A reference to the leader if known, null otherwise
 602      */
 603     public ActorSelection getLeader(){
 604         String leaderAddress = getLeaderAddress();
 605
 606         if(leaderAddress == null){
 607             return null;
 608         }
 609
 610         return context.actorSelection(leaderAddress);
 611     }
 612
 613     /**
 614      *
 615      * @return the current leader's id
 616      */
 617     protected final String getLeaderId(){
 618         return getCurrentBehavior().getLeaderId();
 619     }
 620
 621     @VisibleForTesting
 622     protected final RaftState getRaftState() {
 623         return getCurrentBehavior().state();
 624     }
 625
 626     protected Long getCurrentTerm(){
 627         return context.getTermInformation().getCurrentTerm();
 628     }
 629
 630     protected RaftActorContext getRaftActorContext() {
 631         return context;
 632     }
 633
 634     protected void updateConfigParams(ConfigParams configParams) {
 635
 636         // obtain the RaftPolicy for oldConfigParams and the updated one.
 637         String oldRaftPolicy = context.getConfigParams().
 638             getCustomRaftPolicyImplementationClass();
 639         String newRaftPolicy = configParams.
 640             getCustomRaftPolicyImplementationClass();
 641
 642         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 643             oldRaftPolicy, newRaftPolicy);
 644         context.setConfigParams(configParams);
 645         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 646             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 647             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 648             // avoids potential disruption. Otherwise, switch to Follower normally.
 649             RaftActorBehavior behavior = getCurrentBehavior();
 650             if (behavior != null && behavior.state() == RaftState.Follower) {
 651                 String previousLeaderId = behavior.getLeaderId();
 652                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 653
 654                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
 655
 656                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 657             } else {
 658                 initializeBehavior();
 659             }
 660         }
 661     }
 662
 663     public final DataPersistenceProvider persistence() {
 664         return delegatingPersistenceProvider.getDelegate();
 665     }
 666
 667     public void setPersistence(DataPersistenceProvider provider) {
 668         delegatingPersistenceProvider.setDelegate(provider);
 669     }
 670
 671     protected void setPersistence(boolean persistent) {
 672         DataPersistenceProvider currentPersistence = persistence();
 673         if(persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
 674             setPersistence(new PersistentDataProvider(this));
 675
 676             if(getCurrentBehavior() != null) {
 677                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
 678                 captureSnapshot();
 679             }
 680         } else if(!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
 681             setPersistence(new NonPersistentDataProvider() {
 682                 /**
 683                  * The way snapshotting works is,
 684                  * <ol>
 685                  * <li> RaftActor calls createSnapshot on the Shard
 686                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 687                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 688                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 689                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 690                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 691                  * </ol>
 692                  */
 693                 @Override
 694                 public void saveSnapshot(Object o) {
 695                     // Make saving Snapshot successful
 696                     // Committing the snapshot here would end up calling commit in the creating state which would
 697                     // be a state violation. That's why now we send a message to commit the snapshot.
 698                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 699                 }
 700             });
 701         }
 702     }
 703
 704     /**
 705      * setPeerAddress sets the address of a known peer at a later time.
 706      * <p>
 707      * This is to account for situations where a we know that a peer
 708      * exists but we do not know an address up-front. This may also be used in
 709      * situations where a known peer starts off in a different location and we
 710      * need to change it's address
 711      * <p>
 712      * Note that if the peerId does not match the list of peers passed to
 713      * this actor during construction an IllegalStateException will be thrown.
 714      *
 715      * @param peerId
 716      * @param peerAddress
 717      */
 718     protected void setPeerAddress(String peerId, String peerAddress){
 719         context.setPeerAddress(peerId, peerAddress);
 720     }
 721
 722     /**
 723      * The applyState method will be called by the RaftActor when some data
 724      * needs to be applied to the actor's state
 725      *
 726      * @param clientActor A reference to the client who sent this message. This
 727      *                    is the same reference that was passed to persistData
 728      *                    by the derived actor. clientActor may be null when
 729      *                    the RaftActor is behaving as a follower or during
 730      *                    recovery.
 731      * @param identifier  The identifier of the persisted data. This is also
 732      *                    the same identifier that was passed to persistData by
 733      *                    the derived actor. identifier may be null when
 734      *                    the RaftActor is behaving as a follower or during
 735      *                    recovery
 736      * @param data        A piece of data that was persisted by the persistData call.
 737      *                    This should NEVER be null.
 738      */
 739     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 740
 741     /**
 742      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 743      */
 744     @Nonnull
 745     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 746
 747     /**
 748      * This method is called when recovery is complete.
 749      */
 750     protected abstract void onRecoveryComplete();
 751
 752     /**
 753      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
 754      */
 755     @Nonnull
 756     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 757
 758     /**
 759      * This method will be called by the RaftActor when the state of the
 760      * RaftActor changes. The derived actor can then use methods like
 761      * isLeader or getLeader to do something useful
 762      */
 763     protected abstract void onStateChanged();
 764
 765     /**
 766      * Notifier Actor for this RaftActor to notify when a role change happens
 767      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 768      */
 769     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 770
 771     /**
 772      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 773      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 774      * work prior to performing the operation. On completion of any work, the run method must be called on the
 775      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 776      * this actor's thread dispatcher as as it modifies internal state.
 777      * <p>
 778      * The default implementation immediately runs the operation.
 779      *
 780      * @param operation the operation to run
 781      */
 782     protected void pauseLeader(Runnable operation) {
 783         operation.run();
 784     }
 785
 786     protected void onLeaderChanged(String oldLeader, String newLeader) {
 787
 788     };
 789
 790     private String getLeaderAddress(){
 791         if(isLeader()){
 792             return getSelf().path().toString();
 793         }
 794         String leaderId = getLeaderId();
 795         if (leaderId == null) {
 796             return null;
 797         }
 798         String peerAddress = context.getPeerAddress(leaderId);
 799         if(LOG.isDebugEnabled()) {
 800             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 801                     persistenceId(), leaderId, peerAddress);
 802         }
 803
 804         return peerAddress;
 805     }
 806
 807     protected boolean hasFollowers(){
 808         return getRaftActorContext().hasFollowers();
 809     }
 810
 811     private void captureSnapshot() {
 812         SnapshotManager snapshotManager = context.getSnapshotManager();
 813
 814         if (!snapshotManager.isCapturing()) {
 815             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 816             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 817                 replicatedLog().last(), idx);
 818
 819             snapshotManager.capture(replicatedLog().last(), idx);
 820         }
 821     }
 822
 823     /**
 824      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 825      * in which case we need to step down.
 826      */
 827     void becomeNonVoting() {
 828         if (isLeader()) {
 829             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 830                 @Override
 831                 public void onSuccess(ActorRef raftActorRef) {
 832                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 833                     ensureFollowerState();
 834                 }
 835
 836                 @Override
 837                 public void onFailure(ActorRef raftActorRef) {
 838                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 839                     ensureFollowerState();
 840                 }
 841
 842                 private void ensureFollowerState() {
 843                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 844                     // switch to Follower so ensure that.
 845                     if (getRaftState() != RaftState.Follower) {
 846                         initializeBehavior();
 847                     }
 848                 }
 849             });
 850         }
 851     }
 852
 853     /**
 854      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 855      */
 856     private static abstract class BehaviorState implements Immutable {
 857         @Nullable abstract RaftActorBehavior getBehavior();
 858         @Nullable abstract String getLastValidLeaderId();
 859         @Nullable abstract String getLastLeaderId();
 860         @Nullable abstract short getLeaderPayloadVersion();
 861     }
 862
 863     /**
 864      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
 865      */
 866     private static final class SimpleBehaviorState extends BehaviorState {
 867         private final RaftActorBehavior behavior;
 868         private final String lastValidLeaderId;
 869         private final String lastLeaderId;
 870         private final short leaderPayloadVersion;
 871
 872         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
 873                 final RaftActorBehavior behavior) {
 874             this.lastValidLeaderId = lastValidLeaderId;
 875             this.lastLeaderId = lastLeaderId;
 876             this.behavior = Preconditions.checkNotNull(behavior);
 877             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 878         }
 879
 880         @Override
 881         RaftActorBehavior getBehavior() {
 882             return behavior;
 883         }
 884
 885         @Override
 886         String getLastValidLeaderId() {
 887             return lastValidLeaderId;
 888         }
 889
 890         @Override
 891         short getLeaderPayloadVersion() {
 892             return leaderPayloadVersion;
 893         }
 894
 895         @Override
 896         String getLastLeaderId() {
 897             return lastLeaderId;
 898         }
 899     }
 900
 901     /**
 902      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 903      * An instance is created for each RaftActor. It has two functions:
 904      * - it keeps track of the last leader ID we have encountered since we have been created
 905      * - it creates state capture needed to transition from one behavior to the next
 906      */
 907     private static final class BehaviorStateTracker {
 908         /**
 909          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
 910          * allowed before we receive the first message, we know the leader ID to be null.
 911          */
 912         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
 913             @Override
 914             RaftActorBehavior getBehavior() {
 915                 return null;
 916             }
 917
 918             @Override
 919             String getLastValidLeaderId() {
 920                 return null;
 921             }
 922
 923             @Override
 924             short getLeaderPayloadVersion() {
 925                 return -1;
 926             }
 927
 928             @Override
 929             String getLastLeaderId() {
 930                 return null;
 931             }
 932         };
 933
 934         private String lastValidLeaderId;
 935         private String lastLeaderId;
 936
 937         BehaviorState capture(final RaftActorBehavior behavior) {
 938             if (behavior == null) {
 939                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
 940                 return NULL_BEHAVIOR_STATE;
 941             }
 942
 943             lastLeaderId = behavior.getLeaderId();
 944             if (lastLeaderId != null) {
 945                 lastValidLeaderId = lastLeaderId;
 946             }
 947
 948             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
 949         }
 950     }
 951
 952 }