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.ApplyState;
  38 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
  39 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
  40 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  41 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
  42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
  43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  44 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  45 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  48 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
  49 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
  50 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
  51 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
  52 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
  53 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
  54 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  55 import org.opendaylight.yangtools.concepts.Identifier;
  56 import org.opendaylight.yangtools.concepts.Immutable;
  57 import org.slf4j.Logger;
  58 import org.slf4j.LoggerFactory;
  59
  60 /**
  61  * RaftActor encapsulates a state machine that needs to be kept synchronized
  62  * in a cluster. It implements the RAFT algorithm as described in the paper
  63  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  64  * In Search of an Understandable Consensus Algorithm</a>
  65  * <p/>
  66  * RaftActor has 3 states and each state has a certain behavior associated
  67  * with it. A Raft actor can behave as,
  68  * <ul>
  69  * <li> A Leader </li>
  70  * <li> A Follower (or) </li>
  71  * <li> A Candidate </li>
  72  * </ul>
  73  * <p/>
  74  * <p/>
  75  * A RaftActor MUST be a Leader in order to accept requests from clients to
  76  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  77  * a Leader it is also responsible for ensuring that all followers ultimately
  78  * have the same log and therefore the same state machine as itself.
  79  * <p/>
  80  * <p/>
  81  * The current behavior of a RaftActor determines how election for leadership
  82  * is initiated and how peer RaftActors react to request for votes.
  83  * <p/>
  84  * <p/>
  85  * Each RaftActor also needs to know the current election term. It uses this
  86  * information for a couple of things. One is to simply figure out who it
  87  * voted for in the last election. Another is to figure out if the message
  88  * it received to update it's state is stale.
  89  * <p/>
  90  * <p/>
  91  * The RaftActor uses akka-persistence to store it's replicated log.
  92  * Furthermore through it's behaviors a Raft Actor determines
  93  * <p/>
  94  * <ul>
  95  * <li> when a log entry should be persisted </li>
  96  * <li> when a log entry should be applied to the state machine (and) </li>
  97  * <li> when a snapshot should be saved </li>
  98  * </ul>
  99  */
 100 public abstract class RaftActor extends AbstractUntypedPersistentActor {
 101
 102     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
 103
 104     protected final Logger LOG = LoggerFactory.getLogger(getClass());
 105
 106     /**
 107      * This context should NOT be passed directly to any other actor it is
 108      * only to be consumed by the RaftActorBehaviors
 109      */
 110     private final RaftActorContextImpl context;
 111
 112     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
 113
 114     private final PersistentDataProvider persistentProvider;
 115
 116     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
 117
 118     private RaftActorRecoverySupport raftRecovery;
 119
 120     private RaftActorSnapshotMessageSupport snapshotSupport;
 121
 122     private RaftActorServerConfigurationSupport serverConfigurationSupport;
 123
 124     private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
 125
 126     private boolean shuttingDown;
 127
 128     public RaftActor(String id, Map<String, String> peerAddresses,
 129          Optional<ConfigParams> configParams, short payloadVersion) {
 130
 131         persistentProvider = new PersistentDataProvider(this);
 132         delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
 133
 134         context = new RaftActorContextImpl(this.getSelf(),
 135             this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
 136             -1, -1, peerAddresses,
 137             configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl(),
 138             delegatingPersistenceProvider, LOG);
 139
 140         context.setPayloadVersion(payloadVersion);
 141         context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
 142     }
 143
 144     @Override
 145     public void preStart() throws Exception {
 146         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 147                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 148
 149         super.preStart();
 150
 151         snapshotSupport = newRaftActorSnapshotMessageSupport();
 152         serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
 153     }
 154
 155     @Override
 156     public void postStop() {
 157         context.close();
 158         super.postStop();
 159     }
 160
 161     @Override
 162     protected void handleRecover(Object message) {
 163         if(raftRecovery == null) {
 164             raftRecovery = newRaftActorRecoverySupport();
 165         }
 166
 167         boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
 168         if(recoveryComplete) {
 169             onRecoveryComplete();
 170
 171             initializeBehavior();
 172
 173             raftRecovery = null;
 174         }
 175     }
 176
 177     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
 178         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
 179     }
 180
 181     @VisibleForTesting
 182     void initializeBehavior(){
 183         changeCurrentBehavior(new Follower(context));
 184     }
 185
 186     @VisibleForTesting
 187     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
 188         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 189         if (currentBehavior != null) {
 190             try {
 191                 currentBehavior.close();
 192             } catch (Exception e) {
 193                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
 194             }
 195         }
 196
 197         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 198         setCurrentBehavior(newBehavior);
 199         handleBehaviorChange(state, newBehavior);
 200     }
 201
 202     /**
 203      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
 204      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
 205      * implementation for messages which they do not handle
 206      *
 207      * @param message Incoming command message
 208      */
 209     protected void handleNonRaftCommand(final Object message) {
 210         unhandled(message);
 211     }
 212
 213     /**
 214      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
 215      * {@link #handleNonRaftCommand(Object)} instead.
 216      */
 217     @Deprecated
 218     @Override
 219     // FIXME: make this method final once our unit tests do not need to override it
 220     protected void handleCommand(final Object message) {
 221         if (serverConfigurationSupport.handleMessage(message, getSender())) {
 222             return;
 223         }
 224         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
 225             return;
 226         }
 227
 228         if (message instanceof ApplyState) {
 229             ApplyState applyState = (ApplyState) message;
 230
 231             long startTime = System.nanoTime();
 232
 233             if(LOG.isDebugEnabled()) {
 234                 LOG.debug("{}: Applying state for log index {} data {}",
 235                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 236                     applyState.getReplicatedLogEntry().getData());
 237             }
 238
 239             if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
 240                 applyState(applyState.getClientActor(), applyState.getIdentifier(),
 241                     applyState.getReplicatedLogEntry().getData());
 242             }
 243
 244             long elapsedTime = System.nanoTime() - startTime;
 245             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 246                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 247                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 248             }
 249
 250             if (!hasFollowers()) {
 251                 // for single node, the capture should happen after the apply state
 252                 // as we delete messages from the persistent journal which have made it to the snapshot
 253                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
 254                 // and recovery shows data missing
 255                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
 256
 257                 context.getSnapshotManager().trimLog(context.getLastApplied());
 258             }
 259
 260             // Send it to the current behavior - some behaviors like PreLeader need to be notified of ApplyState.
 261             possiblyHandleBehaviorMessage(message);
 262
 263         } else if (message instanceof ApplyJournalEntries) {
 264             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 265             if(LOG.isDebugEnabled()) {
 266                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
 267             }
 268
 269             persistence().persist(applyEntries, NoopProcedure.instance());
 270
 271         } else if (message instanceof FindLeader) {
 272             getSender().tell(
 273                 new FindLeaderReply(getLeaderAddress()),
 274                 getSelf()
 275             );
 276         } else if(message instanceof GetOnDemandRaftState) {
 277             onGetOnDemandRaftStats();
 278         } else if(message instanceof InitiateCaptureSnapshot) {
 279             captureSnapshot();
 280         } else if(message instanceof SwitchBehavior) {
 281             switchBehavior((SwitchBehavior) message);
 282         } else if(message instanceof LeaderTransitioning) {
 283             onLeaderTransitioning();
 284         } else if(message instanceof Shutdown) {
 285             onShutDown();
 286         } else if(message instanceof Runnable) {
 287             ((Runnable)message).run();
 288         } else if(message instanceof NoopPayload) {
 289             persistData(null, null, (NoopPayload)message);
 290         } else if (!possiblyHandleBehaviorMessage(message)) {
 291             handleNonRaftCommand(message);
 292         }
 293     }
 294
 295     private boolean possiblyHandleBehaviorMessage(final Object message) {
 296         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 297         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 298
 299         // A behavior indicates that it processed the change by returning a reference to the next behavior
 300         // to be used. A null return indicates it has not processed the message and we should be passing it to
 301         // the subclass for handling.
 302         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 303         if (nextBehavior != null) {
 304             switchBehavior(state, nextBehavior);
 305             return true;
 306         }
 307
 308         return false;
 309     }
 310
 311     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
 312         LOG.debug("{}: Initiating leader transfer", persistenceId());
 313
 314         if(leadershipTransferInProgress == null) {
 315             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
 316             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 317                 @Override
 318                 public void onSuccess(ActorRef raftActorRef) {
 319                     leadershipTransferInProgress = null;
 320                 }
 321
 322                 @Override
 323                 public void onFailure(ActorRef raftActorRef) {
 324                     leadershipTransferInProgress = null;
 325                 }
 326             });
 327
 328             leadershipTransferInProgress.addOnComplete(onComplete);
 329             leadershipTransferInProgress.init();
 330         } else {
 331             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 332             leadershipTransferInProgress.addOnComplete(onComplete);
 333         }
 334     }
 335
 336     private void onShutDown() {
 337         LOG.debug("{}: onShutDown", persistenceId());
 338
 339         if(shuttingDown) {
 340             return;
 341         }
 342
 343         shuttingDown = true;
 344
 345         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 346         if (currentBehavior.state() != RaftState.Leader) {
 347             // For non-leaders shutdown is a no-op
 348             self().tell(PoisonPill.getInstance(), self());
 349             return;
 350         }
 351
 352         if (context.hasFollowers()) {
 353             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 354                 @Override
 355                 public void onSuccess(ActorRef raftActorRef) {
 356                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 357                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 358                 }
 359
 360                 @Override
 361                 public void onFailure(ActorRef raftActorRef) {
 362                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 363                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 364                 }
 365             });
 366         } else {
 367             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 368                 @Override
 369                 protected void doRun() {
 370                     self().tell(PoisonPill.getInstance(), self());
 371                 }
 372
 373                 @Override
 374                 protected void doCancel() {
 375                     self().tell(PoisonPill.getInstance(), self());
 376                 }
 377             });
 378         }
 379     }
 380
 381     private void onLeaderTransitioning() {
 382         LOG.debug("{}: onLeaderTransitioning", persistenceId());
 383         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 384         if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
 385             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 386                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 387         }
 388     }
 389
 390     private void switchBehavior(SwitchBehavior message) {
 391         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 392             RaftState newState = message.getNewState();
 393             if( newState == RaftState.Leader || newState == RaftState.Follower) {
 394                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 395                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 396                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 397             } else {
 398                 LOG.warn("Switching to behavior : {} - not supported", newState);
 399             }
 400         }
 401     }
 402
 403     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 404         setCurrentBehavior(nextBehavior);
 405         handleBehaviorChange(oldBehaviorState, nextBehavior);
 406     }
 407
 408     @VisibleForTesting
 409     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 410         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 411     }
 412
 413     private void onGetOnDemandRaftStats() {
 414         // Debugging message to retrieve raft stats.
 415
 416         Map<String, String> peerAddresses = new HashMap<>();
 417         Map<String, Boolean> peerVotingStates = new HashMap<>();
 418         for(PeerInfo info: context.getPeers()) {
 419             peerVotingStates.put(info.getId(), info.isVoting());
 420             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
 421         }
 422
 423         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 424         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 425                 .commitIndex(context.getCommitIndex())
 426                 .currentTerm(context.getTermInformation().getCurrentTerm())
 427                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 428                 .inMemoryJournalLogSize(replicatedLog().size())
 429                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 430                 .lastApplied(context.getLastApplied())
 431                 .lastIndex(replicatedLog().lastIndex())
 432                 .lastTerm(replicatedLog().lastTerm())
 433                 .leader(getLeaderId())
 434                 .raftState(currentBehavior.state().toString())
 435                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 436                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 437                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 438                 .votedFor(context.getTermInformation().getVotedFor())
 439                 .isVoting(context.isVotingMember())
 440                 .peerAddresses(peerAddresses)
 441                 .peerVotingStates(peerVotingStates)
 442                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 443
 444         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 445         if (lastLogEntry != null) {
 446             builder.lastLogIndex(lastLogEntry.getIndex());
 447             builder.lastLogTerm(lastLogEntry.getTerm());
 448         }
 449
 450         if(getCurrentBehavior() instanceof AbstractLeader) {
 451             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 452             Collection<String> followerIds = leader.getFollowerIds();
 453             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 454             for(String id: followerIds) {
 455                 final FollowerLogInformation info = leader.getFollower(id);
 456                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 457                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
 458                         context.getPeerInfo(info.getId()).isVoting()));
 459             }
 460
 461             builder.followerInfoList(followerInfoList);
 462         }
 463
 464         sender().tell(builder.build(), self());
 465
 466     }
 467
 468     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
 469         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 470
 471         if (oldBehavior != currentBehavior){
 472             onStateChanged();
 473         }
 474
 475         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
 476         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 477         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 478
 479         // it can happen that the state has not changed but the leader has changed.
 480         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 481         if(!Objects.equals(lastLeaderId, currentBehavior.getLeaderId()) ||
 482            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 483             if(roleChangeNotifier.isPresent()) {
 484                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 485                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 486             }
 487
 488             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 489
 490             if(leadershipTransferInProgress != null) {
 491                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 492             }
 493
 494             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 495         }
 496
 497         if (roleChangeNotifier.isPresent() &&
 498                 (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
 499             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 500                     currentBehavior.state().name()), getSelf());
 501         }
 502     }
 503
 504     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 505         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 506     }
 507
 508     @Override
 509     public long snapshotSequenceNr() {
 510         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 511         // so that we can delete the persistent journal based on the saved sequence-number
 512         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
 513         // was saved and not the number we saved.
 514         // We would want to override it , by asking akka to use the last-sequence number known to us.
 515         return context.getSnapshotManager().getLastSequenceNumber();
 516     }
 517
 518     /**
 519      * When a derived RaftActor needs to persist something it must call
 520      * persistData.
 521      *
 522      * @param clientActor
 523      * @param identifier
 524      * @param data
 525      */
 526     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
 527
 528         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 529             context.getReplicatedLog().lastIndex() + 1,
 530             context.getTermInformation().getCurrentTerm(), data);
 531
 532         if(LOG.isDebugEnabled()) {
 533             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 534         }
 535
 536         final RaftActorContext raftContext = getRaftActorContext();
 537
 538         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
 539             if (!hasFollowers()){
 540                 // Increment the Commit Index and the Last Applied values
 541                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
 542                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
 543
 544                 // Apply the state immediately.
 545                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
 546
 547                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 548                 // the state to durable storage
 549                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
 550
 551             } else {
 552                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
 553
 554                 // Send message for replication
 555                 getCurrentBehavior().handleMessage(getSelf(),
 556                         new Replicate(clientActor, identifier, replicatedLogEntry1));
 557             }
 558         });
 559     }
 560
 561     private ReplicatedLog replicatedLog() {
 562         return context.getReplicatedLog();
 563     }
 564
 565     protected String getId() {
 566         return context.getId();
 567     }
 568
 569     @VisibleForTesting
 570     void setCurrentBehavior(RaftActorBehavior behavior) {
 571         context.setCurrentBehavior(behavior);
 572     }
 573
 574     protected RaftActorBehavior getCurrentBehavior() {
 575         return context.getCurrentBehavior();
 576     }
 577
 578     /**
 579      * Derived actors can call the isLeader method to check if the current
 580      * RaftActor is the Leader or not
 581      *
 582      * @return true it this RaftActor is a Leader false otherwise
 583      */
 584     protected boolean isLeader() {
 585         return context.getId().equals(getCurrentBehavior().getLeaderId());
 586     }
 587
 588     protected final boolean isLeaderActive() {
 589         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader &&
 590                 !shuttingDown && !isLeadershipTransferInProgress();
 591     }
 592
 593     private boolean isLeadershipTransferInProgress() {
 594         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 595     }
 596
 597     /**
 598      * Derived actor can call getLeader if they need a reference to the Leader.
 599      * This would be useful for example in forwarding a request to an actor
 600      * which is the leader
 601      *
 602      * @return A reference to the leader if known, null otherwise
 603      */
 604     public ActorSelection getLeader(){
 605         String leaderAddress = getLeaderAddress();
 606
 607         if(leaderAddress == null){
 608             return null;
 609         }
 610
 611         return context.actorSelection(leaderAddress);
 612     }
 613
 614     /**
 615      *
 616      * @return the current leader's id
 617      */
 618     protected final String getLeaderId(){
 619         return getCurrentBehavior().getLeaderId();
 620     }
 621
 622     @VisibleForTesting
 623     protected final RaftState getRaftState() {
 624         return getCurrentBehavior().state();
 625     }
 626
 627     protected Long getCurrentTerm(){
 628         return context.getTermInformation().getCurrentTerm();
 629     }
 630
 631     protected RaftActorContext getRaftActorContext() {
 632         return context;
 633     }
 634
 635     protected void updateConfigParams(ConfigParams configParams) {
 636
 637         // obtain the RaftPolicy for oldConfigParams and the updated one.
 638         String oldRaftPolicy = context.getConfigParams().
 639             getCustomRaftPolicyImplementationClass();
 640         String newRaftPolicy = configParams.
 641             getCustomRaftPolicyImplementationClass();
 642
 643         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 644             oldRaftPolicy, newRaftPolicy);
 645         context.setConfigParams(configParams);
 646         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 647             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 648             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 649             // avoids potential disruption. Otherwise, switch to Follower normally.
 650             RaftActorBehavior behavior = getCurrentBehavior();
 651             if (behavior != null && behavior.state() == RaftState.Follower) {
 652                 String previousLeaderId = behavior.getLeaderId();
 653                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 654
 655                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
 656
 657                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 658             } else {
 659                 initializeBehavior();
 660             }
 661         }
 662     }
 663
 664     public final DataPersistenceProvider persistence() {
 665         return delegatingPersistenceProvider.getDelegate();
 666     }
 667
 668     public void setPersistence(DataPersistenceProvider provider) {
 669         delegatingPersistenceProvider.setDelegate(provider);
 670     }
 671
 672     protected void setPersistence(boolean persistent) {
 673         DataPersistenceProvider currentPersistence = persistence();
 674         if(persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
 675             setPersistence(new PersistentDataProvider(this));
 676
 677             if(getCurrentBehavior() != null) {
 678                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
 679                 captureSnapshot();
 680             }
 681         } else if(!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
 682             setPersistence(new NonPersistentDataProvider() {
 683                 /**
 684                  * The way snapshotting works is,
 685                  * <ol>
 686                  * <li> RaftActor calls createSnapshot on the Shard
 687                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 688                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 689                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 690                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 691                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 692                  * </ol>
 693                  */
 694                 @Override
 695                 public void saveSnapshot(Object o) {
 696                     // Make saving Snapshot successful
 697                     // Committing the snapshot here would end up calling commit in the creating state which would
 698                     // be a state violation. That's why now we send a message to commit the snapshot.
 699                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 700                 }
 701             });
 702         }
 703     }
 704
 705     /**
 706      * setPeerAddress sets the address of a known peer at a later time.
 707      * <p>
 708      * This is to account for situations where a we know that a peer
 709      * exists but we do not know an address up-front. This may also be used in
 710      * situations where a known peer starts off in a different location and we
 711      * need to change it's address
 712      * <p>
 713      * Note that if the peerId does not match the list of peers passed to
 714      * this actor during construction an IllegalStateException will be thrown.
 715      *
 716      * @param peerId
 717      * @param peerAddress
 718      */
 719     protected void setPeerAddress(String peerId, String peerAddress){
 720         context.setPeerAddress(peerId, peerAddress);
 721     }
 722
 723     /**
 724      * The applyState method will be called by the RaftActor when some data
 725      * needs to be applied to the actor's state
 726      *
 727      * @param clientActor A reference to the client who sent this message. This
 728      *                    is the same reference that was passed to persistData
 729      *                    by the derived actor. clientActor may be null when
 730      *                    the RaftActor is behaving as a follower or during
 731      *                    recovery.
 732      * @param identifier  The identifier of the persisted data. This is also
 733      *                    the same identifier that was passed to persistData by
 734      *                    the derived actor. identifier may be null when
 735      *                    the RaftActor is behaving as a follower or during
 736      *                    recovery
 737      * @param data        A piece of data that was persisted by the persistData call.
 738      *                    This should NEVER be null.
 739      */
 740     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 741
 742     /**
 743      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 744      */
 745     @Nonnull
 746     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 747
 748     /**
 749      * This method is called when recovery is complete.
 750      */
 751     protected abstract void onRecoveryComplete();
 752
 753     /**
 754      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
 755      */
 756     @Nonnull
 757     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 758
 759     /**
 760      * This method will be called by the RaftActor when the state of the
 761      * RaftActor changes. The derived actor can then use methods like
 762      * isLeader or getLeader to do something useful
 763      */
 764     protected abstract void onStateChanged();
 765
 766     /**
 767      * Notifier Actor for this RaftActor to notify when a role change happens
 768      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 769      */
 770     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 771
 772     /**
 773      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 774      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 775      * work prior to performing the operation. On completion of any work, the run method must be called on the
 776      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 777      * this actor's thread dispatcher as as it modifies internal state.
 778      * <p>
 779      * The default implementation immediately runs the operation.
 780      *
 781      * @param operation the operation to run
 782      */
 783     protected void pauseLeader(Runnable operation) {
 784         operation.run();
 785     }
 786
 787     protected void onLeaderChanged(String oldLeader, String newLeader) {
 788
 789     };
 790
 791     private String getLeaderAddress(){
 792         if(isLeader()){
 793             return getSelf().path().toString();
 794         }
 795         String leaderId = getLeaderId();
 796         if (leaderId == null) {
 797             return null;
 798         }
 799         String peerAddress = context.getPeerAddress(leaderId);
 800         if(LOG.isDebugEnabled()) {
 801             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 802                     persistenceId(), leaderId, peerAddress);
 803         }
 804
 805         return peerAddress;
 806     }
 807
 808     protected boolean hasFollowers(){
 809         return getRaftActorContext().hasFollowers();
 810     }
 811
 812     private void captureSnapshot() {
 813         SnapshotManager snapshotManager = context.getSnapshotManager();
 814
 815         if (!snapshotManager.isCapturing()) {
 816             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 817             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 818                 replicatedLog().last(), idx);
 819
 820             snapshotManager.capture(replicatedLog().last(), idx);
 821         }
 822     }
 823
 824     /**
 825      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 826      * in which case we need to step down.
 827      */
 828     void becomeNonVoting() {
 829         if (isLeader()) {
 830             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 831                 @Override
 832                 public void onSuccess(ActorRef raftActorRef) {
 833                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 834                     ensureFollowerState();
 835                 }
 836
 837                 @Override
 838                 public void onFailure(ActorRef raftActorRef) {
 839                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 840                     ensureFollowerState();
 841                 }
 842
 843                 private void ensureFollowerState() {
 844                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 845                     // switch to Follower so ensure that.
 846                     if (getRaftState() != RaftState.Follower) {
 847                         initializeBehavior();
 848                     }
 849                 }
 850             });
 851         }
 852     }
 853
 854     /**
 855      * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.persisted.DeleteEntries}
 856      *             whose type for fromIndex is long instead of int. This class was kept for backwards
 857      *             compatibility with Helium.
 858      */
 859     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
 860     @SuppressWarnings("serial")
 861     @Deprecated
 862     static class DeleteEntries implements Serializable {
 863         private final int fromIndex;
 864
 865         public DeleteEntries(int fromIndex) {
 866             this.fromIndex = fromIndex;
 867         }
 868
 869         public int getFromIndex() {
 870             return fromIndex;
 871         }
 872
 873         private Object readResolve() {
 874             return org.opendaylight.controller.cluster.raft.persisted.DeleteEntries.createMigrated(fromIndex);
 875         }
 876     }
 877
 878     /**
 879      * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm}
 880      *             which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
 881      */
 882     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
 883     @SuppressWarnings("serial")
 884     @Deprecated
 885     static class UpdateElectionTerm implements Serializable {
 886         private final long currentTerm;
 887         private final String votedFor;
 888
 889         public UpdateElectionTerm(long currentTerm, String votedFor) {
 890             this.currentTerm = currentTerm;
 891             this.votedFor = votedFor;
 892         }
 893
 894         public long getCurrentTerm() {
 895             return currentTerm;
 896         }
 897
 898         public String getVotedFor() {
 899             return votedFor;
 900         }
 901
 902         private Object readResolve() {
 903             return org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm.createMigrated(
 904                     currentTerm, votedFor);
 905         }
 906     }
 907
 908     /**
 909      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 910      */
 911     private static abstract class BehaviorState implements Immutable {
 912         @Nullable abstract RaftActorBehavior getBehavior();
 913         @Nullable abstract String getLastValidLeaderId();
 914         @Nullable abstract String getLastLeaderId();
 915         @Nullable abstract short getLeaderPayloadVersion();
 916     }
 917
 918     /**
 919      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
 920      */
 921     private static final class SimpleBehaviorState extends BehaviorState {
 922         private final RaftActorBehavior behavior;
 923         private final String lastValidLeaderId;
 924         private final String lastLeaderId;
 925         private final short leaderPayloadVersion;
 926
 927         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
 928                 final RaftActorBehavior behavior) {
 929             this.lastValidLeaderId = lastValidLeaderId;
 930             this.lastLeaderId = lastLeaderId;
 931             this.behavior = Preconditions.checkNotNull(behavior);
 932             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 933         }
 934
 935         @Override
 936         RaftActorBehavior getBehavior() {
 937             return behavior;
 938         }
 939
 940         @Override
 941         String getLastValidLeaderId() {
 942             return lastValidLeaderId;
 943         }
 944
 945         @Override
 946         short getLeaderPayloadVersion() {
 947             return leaderPayloadVersion;
 948         }
 949
 950         @Override
 951         String getLastLeaderId() {
 952             return lastLeaderId;
 953         }
 954     }
 955
 956     /**
 957      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 958      * An instance is created for each RaftActor. It has two functions:
 959      * - it keeps track of the last leader ID we have encountered since we have been created
 960      * - it creates state capture needed to transition from one behavior to the next
 961      */
 962     private static final class BehaviorStateTracker {
 963         /**
 964          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
 965          * allowed before we receive the first message, we know the leader ID to be null.
 966          */
 967         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
 968             @Override
 969             RaftActorBehavior getBehavior() {
 970                 return null;
 971             }
 972
 973             @Override
 974             String getLastValidLeaderId() {
 975                 return null;
 976             }
 977
 978             @Override
 979             short getLeaderPayloadVersion() {
 980                 return -1;
 981             }
 982
 983             @Override
 984             String getLastLeaderId() {
 985                 return null;
 986             }
 987         };
 988
 989         private String lastValidLeaderId;
 990         private String lastLeaderId;
 991
 992         BehaviorState capture(final RaftActorBehavior behavior) {
 993             if (behavior == null) {
 994                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
 995                 return NULL_BEHAVIOR_STATE;
 996             }
 997
 998             lastLeaderId = behavior.getLeaderId();
 999             if (lastLeaderId != null) {
1000                 lastValidLeaderId = lastLeaderId;
1001             }
1002
1003             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
1004         }
1005     }
1006
1007 }