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.io.Serializable;
  21 import java.util.Collection;
  22 import java.util.HashMap;
  23 import java.util.List;
  24 import java.util.Map;
  25 import java.util.Objects;
  26 import java.util.concurrent.TimeUnit;
  27 import javax.annotation.Nonnull;
  28 import javax.annotation.Nullable;
  29 import org.apache.commons.lang3.time.DurationFormatUtils;
  30 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  31 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
  32 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
  33 import org.opendaylight.controller.cluster.PersistentDataProvider;
  34 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
  35 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  36 import org.opendaylight.controller.cluster.notifications.RoleChanged;
  37 import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
  38 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  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.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             if (context.getReplicatedLog().size() > 0) {
 175                 self().tell(new InitiateCaptureSnapshot(), self());
 176                 LOG.info("{}: Snapshot capture initiated after recovery", persistenceId());
 177             } else {
 178                 LOG.info("{}: Snapshot capture NOT initiated after recovery, journal empty", persistenceId());
 179             }
 180         }
 181     }
 182
 183     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
 184         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
 185     }
 186
 187     @VisibleForTesting
 188     void initializeBehavior(){
 189         changeCurrentBehavior(new Follower(context));
 190     }
 191
 192     @VisibleForTesting
 193     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
 194         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 195         if (currentBehavior != null) {
 196             try {
 197                 currentBehavior.close();
 198             } catch (Exception e) {
 199                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
 200             }
 201         }
 202
 203         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 204         setCurrentBehavior(newBehavior);
 205         handleBehaviorChange(state, newBehavior);
 206     }
 207
 208     /**
 209      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
 210      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
 211      * implementation for messages which they do not handle
 212      *
 213      * @param message Incoming command message
 214      */
 215     protected void handleNonRaftCommand(final Object message) {
 216         unhandled(message);
 217     }
 218
 219     /**
 220      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
 221      * {@link #handleNonRaftCommand(Object)} instead.
 222      */
 223     @Deprecated
 224     @Override
 225     // FIXME: make this method final once our unit tests do not need to override it
 226     protected void handleCommand(final Object message) {
 227         if (serverConfigurationSupport.handleMessage(message, getSender())) {
 228             return;
 229         }
 230         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
 231             return;
 232         }
 233
 234         if (message instanceof ApplyState) {
 235             ApplyState applyState = (ApplyState) message;
 236
 237             long startTime = System.nanoTime();
 238
 239             if(LOG.isDebugEnabled()) {
 240                 LOG.debug("{}: Applying state for log index {} data {}",
 241                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 242                     applyState.getReplicatedLogEntry().getData());
 243             }
 244
 245             applyState(applyState.getClientActor(), applyState.getIdentifier(),
 246                 applyState.getReplicatedLogEntry().getData());
 247
 248             long elapsedTime = System.nanoTime() - startTime;
 249             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 250                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 251                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 252             }
 253
 254             if (!hasFollowers()) {
 255                 // for single node, the capture should happen after the apply state
 256                 // as we delete messages from the persistent journal which have made it to the snapshot
 257                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
 258                 // and recovery shows data missing
 259                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
 260
 261                 context.getSnapshotManager().trimLog(context.getLastApplied());
 262             }
 263
 264         } else if (message instanceof ApplyJournalEntries) {
 265             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 266             if(LOG.isDebugEnabled()) {
 267                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
 268             }
 269
 270             persistence().persist(applyEntries, NoopProcedure.instance());
 271
 272         } else if (message instanceof FindLeader) {
 273             getSender().tell(
 274                 new FindLeaderReply(getLeaderAddress()),
 275                 getSelf()
 276             );
 277         } else if(message instanceof GetOnDemandRaftState) {
 278             onGetOnDemandRaftStats();
 279         } else if(message instanceof InitiateCaptureSnapshot) {
 280             captureSnapshot();
 281         } else if(message instanceof SwitchBehavior) {
 282             switchBehavior(((SwitchBehavior) message));
 283         } else if(message instanceof LeaderTransitioning) {
 284             onLeaderTransitioning();
 285         } else if(message instanceof Shutdown) {
 286             onShutDown();
 287         } else if(message instanceof Runnable) {
 288             ((Runnable)message).run();
 289         } else {
 290             // Processing the message may affect the state, hence we need to capture it
 291             final RaftActorBehavior currentBehavior = getCurrentBehavior();
 292             final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 293
 294             // A behavior indicates that it processed the change by returning a reference to the next behavior
 295             // to be used. A null return indicates it has not processed the message and we should be passing it to
 296             // the subclass for handling.
 297             final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 298             if (nextBehavior != null) {
 299                 switchBehavior(state, nextBehavior);
 300             } else {
 301                 handleNonRaftCommand(message);
 302             }
 303         }
 304     }
 305
 306     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
 307         LOG.debug("{}: Initiating leader transfer", persistenceId());
 308
 309         if(leadershipTransferInProgress == null) {
 310             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
 311             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 312                 @Override
 313                 public void onSuccess(ActorRef raftActorRef) {
 314                     leadershipTransferInProgress = null;
 315                 }
 316
 317                 @Override
 318                 public void onFailure(ActorRef raftActorRef) {
 319                     leadershipTransferInProgress = null;
 320                 }
 321             });
 322
 323             leadershipTransferInProgress.addOnComplete(onComplete);
 324             leadershipTransferInProgress.init();
 325         } else {
 326             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 327             leadershipTransferInProgress.addOnComplete(onComplete);
 328         }
 329     }
 330
 331     private void onShutDown() {
 332         LOG.debug("{}: onShutDown", persistenceId());
 333
 334         if(shuttingDown) {
 335             return;
 336         }
 337
 338         shuttingDown = true;
 339
 340         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 341         if (currentBehavior.state() != RaftState.Leader) {
 342             // For non-leaders shutdown is a no-op
 343             self().tell(PoisonPill.getInstance(), self());
 344             return;
 345         }
 346
 347         if (context.hasFollowers()) {
 348             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 349                 @Override
 350                 public void onSuccess(ActorRef raftActorRef) {
 351                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 352                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 353                 }
 354
 355                 @Override
 356                 public void onFailure(ActorRef raftActorRef) {
 357                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 358                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 359                 }
 360             });
 361         } else {
 362             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 363                 @Override
 364                 protected void doRun() {
 365                     self().tell(PoisonPill.getInstance(), self());
 366                 }
 367
 368                 @Override
 369                 protected void doCancel() {
 370                     self().tell(PoisonPill.getInstance(), self());
 371                 }
 372             });
 373         }
 374     }
 375
 376     private void onLeaderTransitioning() {
 377         LOG.debug("{}: onLeaderTransitioning", persistenceId());
 378         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 379         if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
 380             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 381                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 382         }
 383     }
 384
 385     private void switchBehavior(SwitchBehavior message) {
 386         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 387             RaftState newState = message.getNewState();
 388             if( newState == RaftState.Leader || newState == RaftState.Follower) {
 389                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 390                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 391                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 392             } else {
 393                 LOG.warn("Switching to behavior : {} - not supported", newState);
 394             }
 395         }
 396     }
 397
 398     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 399         setCurrentBehavior(nextBehavior);
 400         handleBehaviorChange(oldBehaviorState, nextBehavior);
 401     }
 402
 403     @VisibleForTesting
 404     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 405         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 406     }
 407
 408     private void onGetOnDemandRaftStats() {
 409         // Debugging message to retrieve raft stats.
 410
 411         Map<String, String> peerAddresses = new HashMap<>();
 412         for(String peerId: context.getPeerIds()) {
 413             peerAddresses.put(peerId, context.getPeerAddress(peerId));
 414         }
 415
 416         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 417         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 418                 .commitIndex(context.getCommitIndex())
 419                 .currentTerm(context.getTermInformation().getCurrentTerm())
 420                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 421                 .inMemoryJournalLogSize(replicatedLog().size())
 422                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 423                 .lastApplied(context.getLastApplied())
 424                 .lastIndex(replicatedLog().lastIndex())
 425                 .lastTerm(replicatedLog().lastTerm())
 426                 .leader(getLeaderId())
 427                 .raftState(currentBehavior.state().toString())
 428                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 429                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 430                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 431                 .votedFor(context.getTermInformation().getVotedFor())
 432                 .peerAddresses(peerAddresses)
 433                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 434
 435         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 436         if (lastLogEntry != null) {
 437             builder.lastLogIndex(lastLogEntry.getIndex());
 438             builder.lastLogTerm(lastLogEntry.getTerm());
 439         }
 440
 441         if(getCurrentBehavior() instanceof AbstractLeader) {
 442             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 443             Collection<String> followerIds = leader.getFollowerIds();
 444             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 445             for(String id: followerIds) {
 446                 final FollowerLogInformation info = leader.getFollower(id);
 447                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 448                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity())));
 449             }
 450
 451             builder.followerInfoList(followerInfoList);
 452         }
 453
 454         sender().tell(builder.build(), self());
 455
 456     }
 457
 458     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
 459         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 460
 461         if (oldBehavior != currentBehavior){
 462             onStateChanged();
 463         }
 464
 465         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 466         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 467
 468         // it can happen that the state has not changed but the leader has changed.
 469         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 470         if(!Objects.equals(lastValidLeaderId, currentBehavior.getLeaderId()) ||
 471            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 472             if(roleChangeNotifier.isPresent()) {
 473                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 474                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 475             }
 476
 477             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 478
 479             if(leadershipTransferInProgress != null) {
 480                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 481             }
 482
 483             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 484         }
 485
 486         if (roleChangeNotifier.isPresent() &&
 487                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
 488             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 489                     currentBehavior.state().name()), getSelf());
 490         }
 491     }
 492
 493     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 494         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 495     }
 496
 497     @Override
 498     public long snapshotSequenceNr() {
 499         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 500         // so that we can delete the persistent journal based on the saved sequence-number
 501         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
 502         // was saved and not the number we saved.
 503         // We would want to override it , by asking akka to use the last-sequence number known to us.
 504         return context.getSnapshotManager().getLastSequenceNumber();
 505     }
 506
 507     /**
 508      * When a derived RaftActor needs to persist something it must call
 509      * persistData.
 510      *
 511      * @param clientActor
 512      * @param identifier
 513      * @param data
 514      */
 515     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
 516
 517         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 518             context.getReplicatedLog().lastIndex() + 1,
 519             context.getTermInformation().getCurrentTerm(), data);
 520
 521         if(LOG.isDebugEnabled()) {
 522             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 523         }
 524
 525         final RaftActorContext raftContext = getRaftActorContext();
 526
 527         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
 528             if (!hasFollowers()){
 529                 // Increment the Commit Index and the Last Applied values
 530                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
 531                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
 532
 533                 // Apply the state immediately.
 534                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
 535
 536                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 537                 // the state to durable storage
 538                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
 539
 540             } else if (clientActor != null) {
 541                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
 542
 543                 // Send message for replication
 544                 getCurrentBehavior().handleMessage(getSelf(),
 545                         new Replicate(clientActor, identifier, replicatedLogEntry1));
 546             }
 547         });
 548     }
 549
 550     private ReplicatedLog replicatedLog() {
 551         return context.getReplicatedLog();
 552     }
 553
 554     protected String getId() {
 555         return context.getId();
 556     }
 557
 558     @VisibleForTesting
 559     void setCurrentBehavior(RaftActorBehavior behavior) {
 560         context.setCurrentBehavior(behavior);
 561     }
 562
 563     protected RaftActorBehavior getCurrentBehavior() {
 564         return context.getCurrentBehavior();
 565     }
 566
 567     /**
 568      * Derived actors can call the isLeader method to check if the current
 569      * RaftActor is the Leader or not
 570      *
 571      * @return true it this RaftActor is a Leader false otherwise
 572      */
 573     protected boolean isLeader() {
 574         return context.getId().equals(getCurrentBehavior().getLeaderId());
 575     }
 576
 577     protected final boolean isLeaderActive() {
 578         return getRaftState() != RaftState.IsolatedLeader && !shuttingDown &&
 579                 !isLeadershipTransferInProgress();
 580     }
 581
 582     private boolean isLeadershipTransferInProgress() {
 583         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 584     }
 585
 586     /**
 587      * Derived actor can call getLeader if they need a reference to the Leader.
 588      * This would be useful for example in forwarding a request to an actor
 589      * which is the leader
 590      *
 591      * @return A reference to the leader if known, null otherwise
 592      */
 593     protected ActorSelection getLeader(){
 594         String leaderAddress = getLeaderAddress();
 595
 596         if(leaderAddress == null){
 597             return null;
 598         }
 599
 600         return context.actorSelection(leaderAddress);
 601     }
 602
 603     /**
 604      *
 605      * @return the current leader's id
 606      */
 607     protected final String getLeaderId(){
 608         return getCurrentBehavior().getLeaderId();
 609     }
 610
 611     @VisibleForTesting
 612     protected final RaftState getRaftState() {
 613         return getCurrentBehavior().state();
 614     }
 615
 616     protected Long getCurrentTerm(){
 617         return context.getTermInformation().getCurrentTerm();
 618     }
 619
 620     protected RaftActorContext getRaftActorContext() {
 621         return context;
 622     }
 623
 624     protected void updateConfigParams(ConfigParams configParams) {
 625
 626         // obtain the RaftPolicy for oldConfigParams and the updated one.
 627         String oldRaftPolicy = context.getConfigParams().
 628             getCustomRaftPolicyImplementationClass();
 629         String newRaftPolicy = configParams.
 630             getCustomRaftPolicyImplementationClass();
 631
 632         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 633             oldRaftPolicy, newRaftPolicy);
 634         context.setConfigParams(configParams);
 635         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 636             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 637             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 638             // avoids potential disruption. Otherwise, switch to Follower normally.
 639             RaftActorBehavior behavior = getCurrentBehavior();
 640             if (behavior != null && behavior.state() == RaftState.Follower) {
 641                 String previousLeaderId = behavior.getLeaderId();
 642                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 643
 644                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
 645
 646                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 647             } else {
 648                 initializeBehavior();
 649             }
 650         }
 651     }
 652
 653     public final DataPersistenceProvider persistence() {
 654         return delegatingPersistenceProvider.getDelegate();
 655     }
 656
 657     public void setPersistence(DataPersistenceProvider provider) {
 658         delegatingPersistenceProvider.setDelegate(provider);
 659     }
 660
 661     protected void setPersistence(boolean persistent) {
 662         if(persistent) {
 663             setPersistence(new PersistentDataProvider(this));
 664         } else {
 665             setPersistence(new NonPersistentDataProvider() {
 666                 /**
 667                  * The way snapshotting works is,
 668                  * <ol>
 669                  * <li> RaftActor calls createSnapshot on the Shard
 670                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 671                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 672                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 673                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 674                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 675                  * </ol>
 676                  */
 677                 @Override
 678                 public void saveSnapshot(Object o) {
 679                     // Make saving Snapshot successful
 680                     // Committing the snapshot here would end up calling commit in the creating state which would
 681                     // be a state violation. That's why now we send a message to commit the snapshot.
 682                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 683                 }
 684             });
 685         }
 686     }
 687
 688     /**
 689      * setPeerAddress sets the address of a known peer at a later time.
 690      * <p>
 691      * This is to account for situations where a we know that a peer
 692      * exists but we do not know an address up-front. This may also be used in
 693      * situations where a known peer starts off in a different location and we
 694      * need to change it's address
 695      * <p>
 696      * Note that if the peerId does not match the list of peers passed to
 697      * this actor during construction an IllegalStateException will be thrown.
 698      *
 699      * @param peerId
 700      * @param peerAddress
 701      */
 702     protected void setPeerAddress(String peerId, String peerAddress){
 703         context.setPeerAddress(peerId, peerAddress);
 704     }
 705
 706     /**
 707      * The applyState method will be called by the RaftActor when some data
 708      * needs to be applied to the actor's state
 709      *
 710      * @param clientActor A reference to the client who sent this message. This
 711      *                    is the same reference that was passed to persistData
 712      *                    by the derived actor. clientActor may be null when
 713      *                    the RaftActor is behaving as a follower or during
 714      *                    recovery.
 715      * @param identifier  The identifier of the persisted data. This is also
 716      *                    the same identifier that was passed to persistData by
 717      *                    the derived actor. identifier may be null when
 718      *                    the RaftActor is behaving as a follower or during
 719      *                    recovery
 720      * @param data        A piece of data that was persisted by the persistData call.
 721      *                    This should NEVER be null.
 722      */
 723     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 724
 725     /**
 726      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 727      */
 728     @Nonnull
 729     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 730
 731     /**
 732      * This method is called when recovery is complete.
 733      */
 734     protected abstract void onRecoveryComplete();
 735
 736     /**
 737      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
 738      */
 739     @Nonnull
 740     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 741
 742     /**
 743      * This method will be called by the RaftActor when the state of the
 744      * RaftActor changes. The derived actor can then use methods like
 745      * isLeader or getLeader to do something useful
 746      */
 747     protected abstract void onStateChanged();
 748
 749     /**
 750      * Notifier Actor for this RaftActor to notify when a role change happens
 751      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 752      */
 753     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 754
 755     /**
 756      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 757      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 758      * work prior to performing the operation. On completion of any work, the run method must be called on the
 759      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 760      * this actor's thread dispatcher as as it modifies internal state.
 761      * <p>
 762      * The default implementation immediately runs the operation.
 763      *
 764      * @param operation the operation to run
 765      */
 766     protected void pauseLeader(Runnable operation) {
 767         operation.run();
 768     }
 769
 770     protected void onLeaderChanged(String oldLeader, String newLeader) {
 771
 772     };
 773
 774     private String getLeaderAddress(){
 775         if(isLeader()){
 776             return getSelf().path().toString();
 777         }
 778         String leaderId = getLeaderId();
 779         if (leaderId == null) {
 780             return null;
 781         }
 782         String peerAddress = context.getPeerAddress(leaderId);
 783         if(LOG.isDebugEnabled()) {
 784             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 785                     persistenceId(), leaderId, peerAddress);
 786         }
 787
 788         return peerAddress;
 789     }
 790
 791     protected boolean hasFollowers(){
 792         return getRaftActorContext().hasFollowers();
 793     }
 794
 795     private void captureSnapshot() {
 796         SnapshotManager snapshotManager = context.getSnapshotManager();
 797
 798         if (!snapshotManager.isCapturing()) {
 799             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 800             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 801                 replicatedLog().last(), idx);
 802
 803             snapshotManager.capture(replicatedLog().last(), idx);
 804         }
 805     }
 806
 807     /**
 808      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 809      * in which case we need to step down.
 810      */
 811     void becomeNonVoting() {
 812         if (isLeader()) {
 813             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 814                 @Override
 815                 public void onSuccess(ActorRef raftActorRef) {
 816                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 817                     ensureFollowerState();
 818                 }
 819
 820                 @Override
 821                 public void onFailure(ActorRef raftActorRef) {
 822                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 823                     ensureFollowerState();
 824                 }
 825
 826                 private void ensureFollowerState() {
 827                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 828                     // switch to Follower so ensure that.
 829                     if (getRaftState() != RaftState.Follower) {
 830                         initializeBehavior();
 831                     }
 832                 }
 833             });
 834         }
 835     }
 836
 837     /**
 838      * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
 839      *             whose type for fromIndex is long instead of int. This class was kept for backwards
 840      *             compatibility with Helium.
 841      */
 842     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
 843     @SuppressWarnings("serial")
 844     @Deprecated
 845     static class DeleteEntries implements Serializable {
 846         private final int fromIndex;
 847
 848         public DeleteEntries(int fromIndex) {
 849             this.fromIndex = fromIndex;
 850         }
 851
 852         public int getFromIndex() {
 853             return fromIndex;
 854         }
 855     }
 856
 857     /**
 858      * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
 859      *             which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
 860      */
 861     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
 862     @SuppressWarnings("serial")
 863     @Deprecated
 864     static class UpdateElectionTerm implements Serializable {
 865         private final long currentTerm;
 866         private final String votedFor;
 867
 868         public UpdateElectionTerm(long currentTerm, String votedFor) {
 869             this.currentTerm = currentTerm;
 870             this.votedFor = votedFor;
 871         }
 872
 873         public long getCurrentTerm() {
 874             return currentTerm;
 875         }
 876
 877         public String getVotedFor() {
 878             return votedFor;
 879         }
 880     }
 881
 882     /**
 883      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 884      */
 885     private static abstract class BehaviorState implements Immutable {
 886         @Nullable abstract RaftActorBehavior getBehavior();
 887         @Nullable abstract String getLastValidLeaderId();
 888         @Nullable abstract short getLeaderPayloadVersion();
 889     }
 890
 891     /**
 892      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
 893      */
 894     private static final class SimpleBehaviorState extends BehaviorState {
 895         private final RaftActorBehavior behavior;
 896         private final String lastValidLeaderId;
 897         private final short leaderPayloadVersion;
 898
 899         SimpleBehaviorState(final String lastValidLeaderId, final RaftActorBehavior behavior) {
 900             this.lastValidLeaderId = lastValidLeaderId;
 901             this.behavior = Preconditions.checkNotNull(behavior);
 902             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 903         }
 904
 905         @Override
 906         RaftActorBehavior getBehavior() {
 907             return behavior;
 908         }
 909
 910         @Override
 911         String getLastValidLeaderId() {
 912             return lastValidLeaderId;
 913         }
 914
 915         @Override
 916         short getLeaderPayloadVersion() {
 917             return leaderPayloadVersion;
 918         }
 919     }
 920
 921     /**
 922      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 923      * An instance is created for each RaftActor. It has two functions:
 924      * - it keeps track of the last leader ID we have encountered since we have been created
 925      * - it creates state capture needed to transition from one behavior to the next
 926      */
 927     private static final class BehaviorStateTracker {
 928         /**
 929          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
 930          * allowed before we receive the first message, we know the leader ID to be null.
 931          */
 932         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
 933             @Override
 934             RaftActorBehavior getBehavior() {
 935                 return null;
 936             }
 937
 938             @Override
 939             String getLastValidLeaderId() {
 940                 return null;
 941             }
 942
 943             @Override
 944             short getLeaderPayloadVersion() {
 945                 return -1;
 946             }
 947         };
 948
 949         private String lastValidLeaderId;
 950
 951         BehaviorState capture(final RaftActorBehavior behavior) {
 952             if (behavior == null) {
 953                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
 954                 return NULL_BEHAVIOR_STATE;
 955             }
 956
 957             final String leaderId = behavior.getLeaderId();
 958             if (leaderId != null) {
 959                 lastValidLeaderId = leaderId;
 960             }
 961
 962             return new SimpleBehaviorState(lastValidLeaderId, behavior);
 963         }
 964     }
 965
 966 }