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     }
 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             applyState(applyState.getClientActor(), applyState.getIdentifier(),
 239                 applyState.getReplicatedLogEntry().getData());
 240
 241             long elapsedTime = System.nanoTime() - startTime;
 242             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 243                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 244                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 245             }
 246
 247             if (!hasFollowers()) {
 248                 // for single node, the capture should happen after the apply state
 249                 // as we delete messages from the persistent journal which have made it to the snapshot
 250                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
 251                 // and recovery shows data missing
 252                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
 253
 254                 context.getSnapshotManager().trimLog(context.getLastApplied());
 255             }
 256
 257         } else if (message instanceof ApplyJournalEntries) {
 258             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 259             if(LOG.isDebugEnabled()) {
 260                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
 261             }
 262
 263             persistence().persist(applyEntries, NoopProcedure.instance());
 264
 265         } else if (message instanceof FindLeader) {
 266             getSender().tell(
 267                 new FindLeaderReply(getLeaderAddress()),
 268                 getSelf()
 269             );
 270         } else if(message instanceof GetOnDemandRaftState) {
 271             onGetOnDemandRaftStats();
 272         } else if(message instanceof InitiateCaptureSnapshot) {
 273             captureSnapshot();
 274         } else if(message instanceof SwitchBehavior) {
 275             switchBehavior(((SwitchBehavior) message));
 276         } else if(message instanceof LeaderTransitioning) {
 277             onLeaderTransitioning();
 278         } else if(message instanceof Shutdown) {
 279             onShutDown();
 280         } else if(message instanceof Runnable) {
 281             ((Runnable)message).run();
 282         } else {
 283             // Processing the message may affect the state, hence we need to capture it
 284             final RaftActorBehavior currentBehavior = getCurrentBehavior();
 285             final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 286
 287             // A behavior indicates that it processed the change by returning a reference to the next behavior
 288             // to be used. A null return indicates it has not processed the message and we should be passing it to
 289             // the subclass for handling.
 290             final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 291             if (nextBehavior != null) {
 292                 switchBehavior(state, nextBehavior);
 293             } else {
 294                 handleNonRaftCommand(message);
 295             }
 296         }
 297     }
 298
 299     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
 300         LOG.debug("{}: Initiating leader transfer", persistenceId());
 301
 302         if(leadershipTransferInProgress == null) {
 303             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
 304             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 305                 @Override
 306                 public void onSuccess(ActorRef raftActorRef) {
 307                     leadershipTransferInProgress = null;
 308                 }
 309
 310                 @Override
 311                 public void onFailure(ActorRef raftActorRef) {
 312                     leadershipTransferInProgress = null;
 313                 }
 314             });
 315
 316             leadershipTransferInProgress.addOnComplete(onComplete);
 317             leadershipTransferInProgress.init();
 318         } else {
 319             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 320             leadershipTransferInProgress.addOnComplete(onComplete);
 321         }
 322     }
 323
 324     private void onShutDown() {
 325         LOG.debug("{}: onShutDown", persistenceId());
 326
 327         if(shuttingDown) {
 328             return;
 329         }
 330
 331         shuttingDown = true;
 332
 333         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 334         if (currentBehavior.state() != RaftState.Leader) {
 335             // For non-leaders shutdown is a no-op
 336             self().tell(PoisonPill.getInstance(), self());
 337             return;
 338         }
 339
 340         if (context.hasFollowers()) {
 341             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 342                 @Override
 343                 public void onSuccess(ActorRef raftActorRef) {
 344                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 345                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 346                 }
 347
 348                 @Override
 349                 public void onFailure(ActorRef raftActorRef) {
 350                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 351                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 352                 }
 353             });
 354         } else {
 355             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 356                 @Override
 357                 protected void doRun() {
 358                     self().tell(PoisonPill.getInstance(), self());
 359                 }
 360
 361                 @Override
 362                 protected void doCancel() {
 363                     self().tell(PoisonPill.getInstance(), self());
 364                 }
 365             });
 366         }
 367     }
 368
 369     private void onLeaderTransitioning() {
 370         LOG.debug("{}: onLeaderTransitioning", persistenceId());
 371         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 372         if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
 373             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 374                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 375         }
 376     }
 377
 378     private void switchBehavior(SwitchBehavior message) {
 379         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 380             RaftState newState = message.getNewState();
 381             if( newState == RaftState.Leader || newState == RaftState.Follower) {
 382                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 383                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 384                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 385             } else {
 386                 LOG.warn("Switching to behavior : {} - not supported", newState);
 387             }
 388         }
 389     }
 390
 391     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 392         setCurrentBehavior(nextBehavior);
 393         handleBehaviorChange(oldBehaviorState, nextBehavior);
 394     }
 395
 396     @VisibleForTesting
 397     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 398         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 399     }
 400
 401     private void onGetOnDemandRaftStats() {
 402         // Debugging message to retrieve raft stats.
 403
 404         Map<String, String> peerAddresses = new HashMap<>();
 405         Map<String, Boolean> peerVotingStates = new HashMap<>();
 406         for(PeerInfo info: context.getPeers()) {
 407             peerVotingStates.put(info.getId(), info.isVoting());
 408             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
 409         }
 410
 411         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 412         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 413                 .commitIndex(context.getCommitIndex())
 414                 .currentTerm(context.getTermInformation().getCurrentTerm())
 415                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 416                 .inMemoryJournalLogSize(replicatedLog().size())
 417                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 418                 .lastApplied(context.getLastApplied())
 419                 .lastIndex(replicatedLog().lastIndex())
 420                 .lastTerm(replicatedLog().lastTerm())
 421                 .leader(getLeaderId())
 422                 .raftState(currentBehavior.state().toString())
 423                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 424                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 425                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 426                 .votedFor(context.getTermInformation().getVotedFor())
 427                 .isVoting(context.isVotingMember())
 428                 .peerAddresses(peerAddresses)
 429                 .peerVotingStates(peerVotingStates)
 430                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 431
 432         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 433         if (lastLogEntry != null) {
 434             builder.lastLogIndex(lastLogEntry.getIndex());
 435             builder.lastLogTerm(lastLogEntry.getTerm());
 436         }
 437
 438         if(getCurrentBehavior() instanceof AbstractLeader) {
 439             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 440             Collection<String> followerIds = leader.getFollowerIds();
 441             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 442             for(String id: followerIds) {
 443                 final FollowerLogInformation info = leader.getFollower(id);
 444                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 445                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
 446                         context.getPeerInfo(info.getId()).isVoting()));
 447             }
 448
 449             builder.followerInfoList(followerInfoList);
 450         }
 451
 452         sender().tell(builder.build(), self());
 453
 454     }
 455
 456     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
 457         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 458
 459         if (oldBehavior != currentBehavior){
 460             onStateChanged();
 461         }
 462
 463         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
 464         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 465         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 466
 467         // it can happen that the state has not changed but the leader has changed.
 468         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 469         if(!Objects.equals(lastLeaderId, currentBehavior.getLeaderId()) ||
 470            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 471             if(roleChangeNotifier.isPresent()) {
 472                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 473                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 474             }
 475
 476             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 477
 478             if(leadershipTransferInProgress != null) {
 479                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 480             }
 481
 482             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 483         }
 484
 485         if (roleChangeNotifier.isPresent() &&
 486                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
 487             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 488                     currentBehavior.state().name()), getSelf());
 489         }
 490     }
 491
 492     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 493         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 494     }
 495
 496     @Override
 497     public long snapshotSequenceNr() {
 498         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 499         // so that we can delete the persistent journal based on the saved sequence-number
 500         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
 501         // was saved and not the number we saved.
 502         // We would want to override it , by asking akka to use the last-sequence number known to us.
 503         return context.getSnapshotManager().getLastSequenceNumber();
 504     }
 505
 506     /**
 507      * When a derived RaftActor needs to persist something it must call
 508      * persistData.
 509      *
 510      * @param clientActor
 511      * @param identifier
 512      * @param data
 513      */
 514     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
 515
 516         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 517             context.getReplicatedLog().lastIndex() + 1,
 518             context.getTermInformation().getCurrentTerm(), data);
 519
 520         if(LOG.isDebugEnabled()) {
 521             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 522         }
 523
 524         final RaftActorContext raftContext = getRaftActorContext();
 525
 526         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
 527             if (!hasFollowers()){
 528                 // Increment the Commit Index and the Last Applied values
 529                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
 530                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
 531
 532                 // Apply the state immediately.
 533                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
 534
 535                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 536                 // the state to durable storage
 537                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
 538
 539             } else if (clientActor != null) {
 540                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
 541
 542                 // Send message for replication
 543                 getCurrentBehavior().handleMessage(getSelf(),
 544                         new Replicate(clientActor, identifier, replicatedLogEntry1));
 545             }
 546         });
 547     }
 548
 549     private ReplicatedLog replicatedLog() {
 550         return context.getReplicatedLog();
 551     }
 552
 553     protected String getId() {
 554         return context.getId();
 555     }
 556
 557     @VisibleForTesting
 558     void setCurrentBehavior(RaftActorBehavior behavior) {
 559         context.setCurrentBehavior(behavior);
 560     }
 561
 562     protected RaftActorBehavior getCurrentBehavior() {
 563         return context.getCurrentBehavior();
 564     }
 565
 566     /**
 567      * Derived actors can call the isLeader method to check if the current
 568      * RaftActor is the Leader or not
 569      *
 570      * @return true it this RaftActor is a Leader false otherwise
 571      */
 572     protected boolean isLeader() {
 573         return context.getId().equals(getCurrentBehavior().getLeaderId());
 574     }
 575
 576     protected final boolean isLeaderActive() {
 577         return getRaftState() != RaftState.IsolatedLeader && !shuttingDown &&
 578                 !isLeadershipTransferInProgress();
 579     }
 580
 581     private boolean isLeadershipTransferInProgress() {
 582         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 583     }
 584
 585     /**
 586      * Derived actor can call getLeader if they need a reference to the Leader.
 587      * This would be useful for example in forwarding a request to an actor
 588      * which is the leader
 589      *
 590      * @return A reference to the leader if known, null otherwise
 591      */
 592     protected ActorSelection getLeader(){
 593         String leaderAddress = getLeaderAddress();
 594
 595         if(leaderAddress == null){
 596             return null;
 597         }
 598
 599         return context.actorSelection(leaderAddress);
 600     }
 601
 602     /**
 603      *
 604      * @return the current leader's id
 605      */
 606     protected final String getLeaderId(){
 607         return getCurrentBehavior().getLeaderId();
 608     }
 609
 610     @VisibleForTesting
 611     protected final RaftState getRaftState() {
 612         return getCurrentBehavior().state();
 613     }
 614
 615     protected Long getCurrentTerm(){
 616         return context.getTermInformation().getCurrentTerm();
 617     }
 618
 619     protected RaftActorContext getRaftActorContext() {
 620         return context;
 621     }
 622
 623     protected void updateConfigParams(ConfigParams configParams) {
 624
 625         // obtain the RaftPolicy for oldConfigParams and the updated one.
 626         String oldRaftPolicy = context.getConfigParams().
 627             getCustomRaftPolicyImplementationClass();
 628         String newRaftPolicy = configParams.
 629             getCustomRaftPolicyImplementationClass();
 630
 631         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 632             oldRaftPolicy, newRaftPolicy);
 633         context.setConfigParams(configParams);
 634         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 635             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 636             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 637             // avoids potential disruption. Otherwise, switch to Follower normally.
 638             RaftActorBehavior behavior = getCurrentBehavior();
 639             if (behavior != null && behavior.state() == RaftState.Follower) {
 640                 String previousLeaderId = behavior.getLeaderId();
 641                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 642
 643                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
 644
 645                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 646             } else {
 647                 initializeBehavior();
 648             }
 649         }
 650     }
 651
 652     public final DataPersistenceProvider persistence() {
 653         return delegatingPersistenceProvider.getDelegate();
 654     }
 655
 656     public void setPersistence(DataPersistenceProvider provider) {
 657         delegatingPersistenceProvider.setDelegate(provider);
 658     }
 659
 660     protected void setPersistence(boolean persistent) {
 661         if(persistent) {
 662             setPersistence(new PersistentDataProvider(this));
 663         } else {
 664             setPersistence(new NonPersistentDataProvider() {
 665                 /**
 666                  * The way snapshotting works is,
 667                  * <ol>
 668                  * <li> RaftActor calls createSnapshot on the Shard
 669                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 670                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 671                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 672                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 673                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 674                  * </ol>
 675                  */
 676                 @Override
 677                 public void saveSnapshot(Object o) {
 678                     // Make saving Snapshot successful
 679                     // Committing the snapshot here would end up calling commit in the creating state which would
 680                     // be a state violation. That's why now we send a message to commit the snapshot.
 681                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 682                 }
 683             });
 684         }
 685     }
 686
 687     /**
 688      * setPeerAddress sets the address of a known peer at a later time.
 689      * <p>
 690      * This is to account for situations where a we know that a peer
 691      * exists but we do not know an address up-front. This may also be used in
 692      * situations where a known peer starts off in a different location and we
 693      * need to change it's address
 694      * <p>
 695      * Note that if the peerId does not match the list of peers passed to
 696      * this actor during construction an IllegalStateException will be thrown.
 697      *
 698      * @param peerId
 699      * @param peerAddress
 700      */
 701     protected void setPeerAddress(String peerId, String peerAddress){
 702         context.setPeerAddress(peerId, peerAddress);
 703     }
 704
 705     /**
 706      * The applyState method will be called by the RaftActor when some data
 707      * needs to be applied to the actor's state
 708      *
 709      * @param clientActor A reference to the client who sent this message. This
 710      *                    is the same reference that was passed to persistData
 711      *                    by the derived actor. clientActor may be null when
 712      *                    the RaftActor is behaving as a follower or during
 713      *                    recovery.
 714      * @param identifier  The identifier of the persisted data. This is also
 715      *                    the same identifier that was passed to persistData by
 716      *                    the derived actor. identifier may be null when
 717      *                    the RaftActor is behaving as a follower or during
 718      *                    recovery
 719      * @param data        A piece of data that was persisted by the persistData call.
 720      *                    This should NEVER be null.
 721      */
 722     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 723
 724     /**
 725      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 726      */
 727     @Nonnull
 728     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 729
 730     /**
 731      * This method is called when recovery is complete.
 732      */
 733     protected abstract void onRecoveryComplete();
 734
 735     /**
 736      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
 737      */
 738     @Nonnull
 739     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 740
 741     /**
 742      * This method will be called by the RaftActor when the state of the
 743      * RaftActor changes. The derived actor can then use methods like
 744      * isLeader or getLeader to do something useful
 745      */
 746     protected abstract void onStateChanged();
 747
 748     /**
 749      * Notifier Actor for this RaftActor to notify when a role change happens
 750      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 751      */
 752     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 753
 754     /**
 755      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 756      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 757      * work prior to performing the operation. On completion of any work, the run method must be called on the
 758      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 759      * this actor's thread dispatcher as as it modifies internal state.
 760      * <p>
 761      * The default implementation immediately runs the operation.
 762      *
 763      * @param operation the operation to run
 764      */
 765     protected void pauseLeader(Runnable operation) {
 766         operation.run();
 767     }
 768
 769     protected void onLeaderChanged(String oldLeader, String newLeader) {
 770
 771     };
 772
 773     private String getLeaderAddress(){
 774         if(isLeader()){
 775             return getSelf().path().toString();
 776         }
 777         String leaderId = getLeaderId();
 778         if (leaderId == null) {
 779             return null;
 780         }
 781         String peerAddress = context.getPeerAddress(leaderId);
 782         if(LOG.isDebugEnabled()) {
 783             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 784                     persistenceId(), leaderId, peerAddress);
 785         }
 786
 787         return peerAddress;
 788     }
 789
 790     protected boolean hasFollowers(){
 791         return getRaftActorContext().hasFollowers();
 792     }
 793
 794     private void captureSnapshot() {
 795         SnapshotManager snapshotManager = context.getSnapshotManager();
 796
 797         if (!snapshotManager.isCapturing()) {
 798             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 799             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 800                 replicatedLog().last(), idx);
 801
 802             snapshotManager.capture(replicatedLog().last(), idx);
 803         }
 804     }
 805
 806     /**
 807      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 808      * in which case we need to step down.
 809      */
 810     void becomeNonVoting() {
 811         if (isLeader()) {
 812             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 813                 @Override
 814                 public void onSuccess(ActorRef raftActorRef) {
 815                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 816                     ensureFollowerState();
 817                 }
 818
 819                 @Override
 820                 public void onFailure(ActorRef raftActorRef) {
 821                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 822                     ensureFollowerState();
 823                 }
 824
 825                 private void ensureFollowerState() {
 826                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 827                     // switch to Follower so ensure that.
 828                     if (getRaftState() != RaftState.Follower) {
 829                         initializeBehavior();
 830                     }
 831                 }
 832             });
 833         }
 834     }
 835
 836     /**
 837      * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
 838      *             whose type for fromIndex is long instead of int. This class was kept for backwards
 839      *             compatibility with Helium.
 840      */
 841     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
 842     @SuppressWarnings("serial")
 843     @Deprecated
 844     static class DeleteEntries implements Serializable {
 845         private final int fromIndex;
 846
 847         public DeleteEntries(int fromIndex) {
 848             this.fromIndex = fromIndex;
 849         }
 850
 851         public int getFromIndex() {
 852             return fromIndex;
 853         }
 854     }
 855
 856     /**
 857      * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
 858      *             which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
 859      */
 860     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
 861     @SuppressWarnings("serial")
 862     @Deprecated
 863     static class UpdateElectionTerm implements Serializable {
 864         private final long currentTerm;
 865         private final String votedFor;
 866
 867         public UpdateElectionTerm(long currentTerm, String votedFor) {
 868             this.currentTerm = currentTerm;
 869             this.votedFor = votedFor;
 870         }
 871
 872         public long getCurrentTerm() {
 873             return currentTerm;
 874         }
 875
 876         public String getVotedFor() {
 877             return votedFor;
 878         }
 879     }
 880
 881     /**
 882      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 883      */
 884     private static abstract class BehaviorState implements Immutable {
 885         @Nullable abstract RaftActorBehavior getBehavior();
 886         @Nullable abstract String getLastValidLeaderId();
 887         @Nullable abstract String getLastLeaderId();
 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 String lastLeaderId;
 898         private final short leaderPayloadVersion;
 899
 900         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
 901                 final RaftActorBehavior behavior) {
 902             this.lastValidLeaderId = lastValidLeaderId;
 903             this.lastLeaderId = lastLeaderId;
 904             this.behavior = Preconditions.checkNotNull(behavior);
 905             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 906         }
 907
 908         @Override
 909         RaftActorBehavior getBehavior() {
 910             return behavior;
 911         }
 912
 913         @Override
 914         String getLastValidLeaderId() {
 915             return lastValidLeaderId;
 916         }
 917
 918         @Override
 919         short getLeaderPayloadVersion() {
 920             return leaderPayloadVersion;
 921         }
 922
 923         @Override
 924         String getLastLeaderId() {
 925             return lastLeaderId;
 926         }
 927     }
 928
 929     /**
 930      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 931      * An instance is created for each RaftActor. It has two functions:
 932      * - it keeps track of the last leader ID we have encountered since we have been created
 933      * - it creates state capture needed to transition from one behavior to the next
 934      */
 935     private static final class BehaviorStateTracker {
 936         /**
 937          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
 938          * allowed before we receive the first message, we know the leader ID to be null.
 939          */
 940         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
 941             @Override
 942             RaftActorBehavior getBehavior() {
 943                 return null;
 944             }
 945
 946             @Override
 947             String getLastValidLeaderId() {
 948                 return null;
 949             }
 950
 951             @Override
 952             short getLeaderPayloadVersion() {
 953                 return -1;
 954             }
 955
 956             @Override
 957             String getLastLeaderId() {
 958                 return null;
 959             }
 960         };
 961
 962         private String lastValidLeaderId;
 963         private String lastLeaderId;
 964
 965         BehaviorState capture(final RaftActorBehavior behavior) {
 966             if (behavior == null) {
 967                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
 968                 return NULL_BEHAVIOR_STATE;
 969             }
 970
 971             lastLeaderId = behavior.getLeaderId();
 972             if (lastLeaderId != null) {
 973                 lastValidLeaderId = lastLeaderId;
 974             }
 975
 976             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
 977         }
 978     }
 979
 980 }