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 package org.opendaylight.controller.cluster.raft;
  10
  11 import static com.google.common.base.Verify.verify;
  12 import static java.util.Objects.requireNonNull;
  13
  14 import akka.actor.ActorRef;
  15 import akka.actor.ActorSelection;
  16 import akka.actor.PoisonPill;
  17 import akka.actor.Status;
  18 import akka.persistence.JournalProtocol;
  19 import akka.persistence.SnapshotProtocol;
  20 import com.google.common.annotations.VisibleForTesting;
  21 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
  22 import java.util.ArrayList;
  23 import java.util.Collection;
  24 import java.util.HashMap;
  25 import java.util.List;
  26 import java.util.Map;
  27 import java.util.Objects;
  28 import java.util.Optional;
  29 import java.util.concurrent.TimeUnit;
  30 import org.apache.commons.lang3.time.DurationFormatUtils;
  31 import org.eclipse.jdt.annotation.NonNull;
  32 import org.eclipse.jdt.annotation.Nullable;
  33 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  34 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
  35 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
  36 import org.opendaylight.controller.cluster.PersistentDataProvider;
  37 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
  38 import org.opendaylight.controller.cluster.mgmt.api.FollowerInfo;
  39 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  40 import org.opendaylight.controller.cluster.notifications.RoleChanged;
  41 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  42 import org.opendaylight.controller.cluster.raft.base.messages.CheckConsensusReached;
  43 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
  44 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
  45 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  46 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
  47 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
  48 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  49 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  50 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  51 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  52 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  53 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
  54 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
  55 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
  56 import org.opendaylight.controller.cluster.raft.messages.Payload;
  57 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
  58 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
  59 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
  60 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
  61 import org.opendaylight.controller.cluster.raft.persisted.SimpleReplicatedLogEntry;
  62 import org.opendaylight.yangtools.concepts.Identifier;
  63 import org.opendaylight.yangtools.concepts.Immutable;
  64
  65 /**
  66  * RaftActor encapsulates a state machine that needs to be kept synchronized
  67  * in a cluster. It implements the RAFT algorithm as described in the paper
  68  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  69  * In Search of an Understandable Consensus Algorithm</a>
  70  *
  71  * <p>
  72  * RaftActor has 3 states and each state has a certain behavior associated
  73  * with it. A Raft actor can behave as,
  74  * <ul>
  75  * <li> A Leader </li>
  76  * <li> A Follower (or) </li>
  77  * <li> A Candidate </li>
  78  * </ul>
  79  *
  80  * <p>
  81  * A RaftActor MUST be a Leader in order to accept requests from clients to
  82  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  83  * a Leader it is also responsible for ensuring that all followers ultimately
  84  * have the same log and therefore the same state machine as itself.
  85  *
  86  * <p>
  87  * The current behavior of a RaftActor determines how election for leadership
  88  * is initiated and how peer RaftActors react to request for votes.
  89  *
  90  * <p>
  91  * Each RaftActor also needs to know the current election term. It uses this
  92  * information for a couple of things. One is to simply figure out who it
  93  * voted for in the last election. Another is to figure out if the message
  94  * it received to update it's state is stale.
  95  *
  96  * <p>
  97  * The RaftActor uses akka-persistence to store it's replicated log.
  98  * Furthermore through it's behaviors a Raft Actor determines
  99  * <ul>
 100  * <li> when a log entry should be persisted </li>
 101  * <li> when a log entry should be applied to the state machine (and) </li>
 102  * <li> when a snapshot should be saved </li>
 103  * </ul>
 104  */
 105 public abstract class RaftActor extends AbstractUntypedPersistentActor {
 106
 107     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
 108
 109     /**
 110      * This context should NOT be passed directly to any other actor it is
 111      * only to be consumed by the RaftActorBehaviors.
 112      */
 113     private final RaftActorContextImpl context;
 114
 115     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
 116
 117     private final PersistentDataProvider persistentProvider;
 118
 119     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
 120
 121     private RaftActorRecoverySupport raftRecovery;
 122
 123     private RaftActorSnapshotMessageSupport snapshotSupport;
 124
 125     private RaftActorServerConfigurationSupport serverConfigurationSupport;
 126
 127     private boolean shuttingDown;
 128
 129     @SuppressFBWarnings(value = "MC_OVERRIDABLE_METHOD_CALL_IN_CONSTRUCTOR", justification = "Akka class design")
 130     protected RaftActor(final String id, final Map<String, String> peerAddresses,
 131          final Optional<ConfigParams> configParams, final short payloadVersion) {
 132
 133         persistentProvider = new PersistentDataProvider(this);
 134         delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
 135
 136         context = new RaftActorContextImpl(getSelf(), getContext(), id,
 137             new ElectionTermImpl(persistentProvider, id, LOG), -1, -1, peerAddresses,
 138             configParams.isPresent() ? configParams.get() : new DefaultConfigParamsImpl(),
 139             delegatingPersistenceProvider, this::handleApplyState, LOG, this::executeInSelf);
 140
 141         context.setPayloadVersion(payloadVersion);
 142         context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
 143     }
 144
 145     @Override
 146     public void preStart() throws Exception {
 147         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 148                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 149
 150         super.preStart();
 151
 152         snapshotSupport = newRaftActorSnapshotMessageSupport();
 153         serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
 154     }
 155
 156     @Override
 157     public void postStop() throws Exception {
 158         context.close();
 159         super.postStop();
 160     }
 161
 162     @Override
 163     protected void handleRecover(final Object message) {
 164         if (raftRecovery == null) {
 165             raftRecovery = newRaftActorRecoverySupport();
 166         }
 167
 168         boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
 169         if (recoveryComplete) {
 170             onRecoveryComplete();
 171
 172             initializeBehavior();
 173
 174             raftRecovery = null;
 175         }
 176     }
 177
 178     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
 179         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
 180     }
 181
 182     @VisibleForTesting
 183     void initializeBehavior() {
 184         changeCurrentBehavior(new Follower(context));
 185     }
 186
 187     @VisibleForTesting
 188     @SuppressWarnings("checkstyle:IllegalCatch")
 189     protected void changeCurrentBehavior(final RaftActorBehavior newBehavior) {
 190         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 191         if (currentBehavior != null) {
 192             try {
 193                 currentBehavior.close();
 194             } catch (Exception e) {
 195                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
 196             }
 197         }
 198
 199         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 200         setCurrentBehavior(newBehavior);
 201         handleBehaviorChange(state, newBehavior);
 202     }
 203
 204     /**
 205      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
 206      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
 207      * implementation for messages which they do not handle
 208      *
 209      * @param message Incoming command message
 210      */
 211     protected void handleNonRaftCommand(final Object message) {
 212         unhandled(message);
 213     }
 214
 215     /**
 216      * Handles a message.
 217      *
 218      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
 219      *             {@link #handleNonRaftCommand(Object)} instead.
 220      */
 221     @Deprecated
 222     @Override
 223     // FIXME: make this method final once our unit tests do not need to override it
 224     protected void handleCommand(final Object message) {
 225         if (serverConfigurationSupport.handleMessage(message, getSender())) {
 226             return;
 227         }
 228         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
 229             return;
 230         }
 231         if (message instanceof ApplyState) {
 232             ApplyState applyState = (ApplyState) message;
 233
 234             if (!hasFollowers()) {
 235                 // for single node, the capture should happen after the apply state
 236                 // as we delete messages from the persistent journal which have made it to the snapshot
 237                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
 238                 // and recovery shows data missing
 239                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
 240
 241                 context.getSnapshotManager().trimLog(context.getLastApplied());
 242             }
 243
 244             possiblyHandleBehaviorMessage(message);
 245         } else if (message instanceof ApplyJournalEntries) {
 246             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 247             LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
 248
 249             persistence().persistAsync(applyEntries, NoopProcedure.instance());
 250         } else if (message instanceof FindLeader) {
 251             getSender().tell(new FindLeaderReply(getLeaderAddress()), getSelf());
 252         } else if (message instanceof GetOnDemandRaftState) {
 253             onGetOnDemandRaftStats();
 254         } else if (message instanceof InitiateCaptureSnapshot) {
 255             captureSnapshot();
 256         } else if (message instanceof SwitchBehavior) {
 257             switchBehavior((SwitchBehavior) message);
 258         } else if (message instanceof LeaderTransitioning) {
 259             onLeaderTransitioning((LeaderTransitioning)message);
 260         } else if (message instanceof Shutdown) {
 261             onShutDown();
 262         } else if (message instanceof Runnable) {
 263             ((Runnable)message).run();
 264         } else if (message instanceof NoopPayload) {
 265             persistData(null, null, (NoopPayload) message, false);
 266         } else if (message instanceof RequestLeadership) {
 267             onRequestLeadership((RequestLeadership) message);
 268         } else if (!possiblyHandleBehaviorMessage(message)) {
 269             if (message instanceof JournalProtocol.Response
 270                 && delegatingPersistenceProvider.handleJournalResponse((JournalProtocol.Response) message)) {
 271                 LOG.debug("{}: handled a journal response", persistenceId());
 272             } else if (message instanceof SnapshotProtocol.Response
 273                 && delegatingPersistenceProvider.handleSnapshotResponse((SnapshotProtocol.Response) message)) {
 274                 LOG.debug("{}: handled a snapshot response", persistenceId());
 275             } else {
 276                 handleNonRaftCommand(message);
 277             }
 278         }
 279     }
 280
 281     private void onRequestLeadership(final RequestLeadership message) {
 282         LOG.debug("{}: onRequestLeadership {}", persistenceId(), message);
 283         if (!isLeader()) {
 284             // non-leader cannot satisfy leadership request
 285             LOG.warn("{}: onRequestLeadership {} was sent to non-leader."
 286                     + " Current behavior: {}. Sending failure response",
 287                     persistenceId(), message, getCurrentBehavior().state());
 288             message.getReplyTo().tell(new LeadershipTransferFailedException("Cannot transfer leader to "
 289                     + message.getRequestedFollowerId()
 290                     + ". RequestLeadership message was sent to non-leader " + persistenceId()), getSelf());
 291             return;
 292         }
 293
 294         final String requestedFollowerId = message.getRequestedFollowerId();
 295         final ActorRef replyTo = message.getReplyTo();
 296         initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 297             @Override
 298             public void onSuccess(final ActorRef raftActorRef) {
 299                 // sanity check
 300                 if (!requestedFollowerId.equals(getLeaderId())) {
 301                     onFailure(raftActorRef);
 302                 }
 303
 304                 LOG.debug("{}: Leadership transferred successfully to {}", persistenceId(), requestedFollowerId);
 305                 replyTo.tell(new Status.Success(null), getSelf());
 306             }
 307
 308             @Override
 309             public void onFailure(final ActorRef raftActorRef) {
 310                 LOG.debug("{}: LeadershipTransfer request from {} failed", persistenceId(), requestedFollowerId);
 311                 replyTo.tell(new Status.Failure(
 312                         new LeadershipTransferFailedException(
 313                                 "Failed to transfer leadership to " + requestedFollowerId
 314                                         + ". Follower is not ready to become leader")),
 315                         getSelf());
 316             }
 317         }, message.getRequestedFollowerId(), RaftActorLeadershipTransferCohort.USE_DEFAULT_LEADER_TIMEOUT);
 318     }
 319
 320     private boolean possiblyHandleBehaviorMessage(final Object message) {
 321         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 322         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 323
 324         // A behavior indicates that it processed the change by returning a reference to the next behavior
 325         // to be used. A null return indicates it has not processed the message and we should be passing it to
 326         // the subclass for handling.
 327         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 328         if (nextBehavior != null) {
 329             switchBehavior(state, nextBehavior);
 330             return true;
 331         }
 332
 333         return false;
 334     }
 335
 336     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete,
 337             final @Nullable String followerId, final long newLeaderTimeoutInMillis) {
 338         LOG.debug("{}: Initiating leader transfer", persistenceId());
 339
 340         RaftActorLeadershipTransferCohort leadershipTransferInProgress = context.getRaftActorLeadershipTransferCohort();
 341         if (leadershipTransferInProgress == null) {
 342             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this, followerId);
 343             leadershipTransferInProgress.setNewLeaderTimeoutInMillis(newLeaderTimeoutInMillis);
 344             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 345                 @Override
 346                 public void onSuccess(final ActorRef raftActorRef) {
 347                     context.setRaftActorLeadershipTransferCohort(null);
 348                 }
 349
 350                 @Override
 351                 public void onFailure(final ActorRef raftActorRef) {
 352                     context.setRaftActorLeadershipTransferCohort(null);
 353                 }
 354             });
 355
 356             leadershipTransferInProgress.addOnComplete(onComplete);
 357
 358             context.setRaftActorLeadershipTransferCohort(leadershipTransferInProgress);
 359             leadershipTransferInProgress.init();
 360
 361         } else {
 362             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 363             leadershipTransferInProgress.addOnComplete(onComplete);
 364         }
 365     }
 366
 367     private void onShutDown() {
 368         LOG.debug("{}: onShutDown", persistenceId());
 369
 370         if (shuttingDown) {
 371             return;
 372         }
 373
 374         shuttingDown = true;
 375
 376         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 377         switch (currentBehavior.state()) {
 378             case Leader:
 379             case PreLeader:
 380                 // Fall-through to more work
 381                 break;
 382             default:
 383                 // For non-leaders shutdown is a no-op
 384                 self().tell(PoisonPill.getInstance(), self());
 385                 return;
 386         }
 387
 388         if (context.hasFollowers()) {
 389             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 390                 @Override
 391                 public void onSuccess(final ActorRef raftActorRef) {
 392                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 393                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 394                 }
 395
 396                 @Override
 397                 public void onFailure(final ActorRef raftActorRef) {
 398                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 399                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 400                 }
 401             }, null, TimeUnit.MILLISECONDS.convert(2, TimeUnit.SECONDS));
 402         } else {
 403             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 404                 @Override
 405                 protected void doRun() {
 406                     self().tell(PoisonPill.getInstance(), self());
 407                 }
 408
 409                 @Override
 410                 protected void doCancel() {
 411                     self().tell(PoisonPill.getInstance(), self());
 412                 }
 413             });
 414         }
 415     }
 416
 417     private void onLeaderTransitioning(final LeaderTransitioning leaderTransitioning) {
 418         LOG.debug("{}: onLeaderTransitioning: {}", persistenceId(), leaderTransitioning);
 419         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 420         if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()
 421                 && leaderTransitioning.getLeaderId().equals(getCurrentBehavior().getLeaderId())) {
 422             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 423                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 424         }
 425     }
 426
 427     private void switchBehavior(final SwitchBehavior message) {
 428         if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 429             RaftState newState = message.getNewState();
 430             if (newState == RaftState.Leader || newState == RaftState.Follower) {
 431                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 432                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 433                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 434             } else {
 435                 LOG.warn("Switching to behavior : {} - not supported", newState);
 436             }
 437         }
 438     }
 439
 440     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 441         setCurrentBehavior(nextBehavior);
 442         handleBehaviorChange(oldBehaviorState, nextBehavior);
 443     }
 444
 445     @VisibleForTesting
 446     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 447         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 448     }
 449
 450     private void onGetOnDemandRaftStats() {
 451         // Debugging message to retrieve raft stats.
 452
 453         Map<String, String> peerAddresses = new HashMap<>();
 454         Map<String, Boolean> peerVotingStates = new HashMap<>();
 455         for (PeerInfo info: context.getPeers()) {
 456             peerVotingStates.put(info.getId(), info.isVoting());
 457             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
 458         }
 459
 460         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 461         OnDemandRaftState.AbstractBuilder<?, ?> builder = newOnDemandRaftStateBuilder()
 462                 .commitIndex(context.getCommitIndex())
 463                 .currentTerm(context.getTermInformation().getCurrentTerm())
 464                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 465                 .inMemoryJournalLogSize(replicatedLog().size())
 466                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 467                 .lastApplied(context.getLastApplied())
 468                 .lastIndex(replicatedLog().lastIndex())
 469                 .lastTerm(replicatedLog().lastTerm())
 470                 .leader(getLeaderId())
 471                 .raftState(currentBehavior.state().toString())
 472                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 473                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 474                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 475                 .votedFor(context.getTermInformation().getVotedFor())
 476                 .isVoting(context.isVotingMember())
 477                 .peerAddresses(peerAddresses)
 478                 .peerVotingStates(peerVotingStates)
 479                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 480
 481         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 482         if (lastLogEntry != null) {
 483             builder.lastLogIndex(lastLogEntry.getIndex());
 484             builder.lastLogTerm(lastLogEntry.getTerm());
 485         }
 486
 487         if (getCurrentBehavior() instanceof AbstractLeader) {
 488             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 489             Collection<String> followerIds = leader.getFollowerIds();
 490             List<FollowerInfo> followerInfoList = new ArrayList<>(followerIds.size());
 491             for (String id: followerIds) {
 492                 final FollowerLogInformation info = leader.getFollower(id);
 493                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 494                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(
 495                             TimeUnit.NANOSECONDS.toMillis(info.nanosSinceLastActivity())),
 496                         context.getPeerInfo(info.getId()).isVoting()));
 497             }
 498
 499             builder.followerInfoList(followerInfoList);
 500         }
 501
 502         sender().tell(builder.build(), self());
 503
 504     }
 505
 506     protected OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
 507         return OnDemandRaftState.builder();
 508     }
 509
 510     private void handleBehaviorChange(final BehaviorState oldBehaviorState, final RaftActorBehavior currentBehavior) {
 511         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 512
 513         if (oldBehavior != currentBehavior) {
 514             onStateChanged();
 515         }
 516
 517         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
 518         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 519         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 520
 521         // it can happen that the state has not changed but the leader has changed.
 522         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 523         if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
 524                 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 525             if (roleChangeNotifier.isPresent()) {
 526                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 527                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 528             }
 529
 530             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 531
 532             RaftActorLeadershipTransferCohort leadershipTransferInProgress =
 533                     context.getRaftActorLeadershipTransferCohort();
 534             if (leadershipTransferInProgress != null) {
 535                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 536             }
 537
 538             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 539         }
 540
 541         if (roleChangeNotifier.isPresent()
 542                 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
 543             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 544                     currentBehavior.state().name()), getSelf());
 545         }
 546     }
 547
 548     private void handleApplyState(final ApplyState applyState) {
 549         long startTime = System.nanoTime();
 550
 551         Payload payload = applyState.getReplicatedLogEntry().getData();
 552         if (LOG.isDebugEnabled()) {
 553             LOG.debug("{}: Applying state for log index {} data {}",
 554                 persistenceId(), applyState.getReplicatedLogEntry().getIndex(), payload);
 555         }
 556
 557         if (!(payload instanceof NoopPayload) && !(payload instanceof ServerConfigurationPayload)) {
 558             applyState(applyState.getClientActor(), applyState.getIdentifier(), payload);
 559         }
 560
 561         long elapsedTime = System.nanoTime() - startTime;
 562         if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
 563             LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 564                     TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 565         }
 566
 567         // Send the ApplyState message back to self to handle further processing asynchronously.
 568         self().tell(applyState, self());
 569     }
 570
 571     protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
 572             final short leaderPayloadVersion) {
 573         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 574     }
 575
 576     @Override
 577     public long snapshotSequenceNr() {
 578         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 579         // so that we can delete the persistent journal based on the saved sequence-number
 580         // However , when akka replays the journal during recovery, it replays it from the sequence number when the
 581         // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
 582         // last-sequence number known to us.
 583         return context.getSnapshotManager().getLastSequenceNumber();
 584     }
 585
 586     /**
 587      * Persists the given Payload in the journal and replicates to any followers. After successful completion,
 588      * {@link #applyState(ActorRef, Identifier, Object)} is notified.
 589      *
 590      * @param clientActor optional ActorRef that is provided via the applyState callback
 591      * @param identifier the payload identifier
 592      * @param data the payload data to persist
 593      * @param batchHint if true, an attempt is made to delay immediate replication and batch the payload with
 594      *        subsequent payloads for efficiency. Otherwise the payload is immediately replicated.
 595      */
 596     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data,
 597             final boolean batchHint) {
 598         ReplicatedLogEntry replicatedLogEntry = new SimpleReplicatedLogEntry(
 599             context.getReplicatedLog().lastIndex() + 1,
 600             context.getTermInformation().getCurrentTerm(), data);
 601         replicatedLogEntry.setPersistencePending(true);
 602
 603         LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 604
 605         final RaftActorContext raftContext = getRaftActorContext();
 606
 607         boolean wasAppended = replicatedLog().appendAndPersist(replicatedLogEntry, persistedLogEntry -> {
 608             // Clear the persistence pending flag in the log entry.
 609             persistedLogEntry.setPersistencePending(false);
 610
 611             if (!hasFollowers()) {
 612                 // Increment the Commit Index and the Last Applied values
 613                 raftContext.setCommitIndex(persistedLogEntry.getIndex());
 614                 raftContext.setLastApplied(persistedLogEntry.getIndex());
 615
 616                 // Apply the state immediately.
 617                 handleApplyState(new ApplyState(clientActor, identifier, persistedLogEntry));
 618
 619                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 620                 // the state to durable storage
 621                 self().tell(new ApplyJournalEntries(persistedLogEntry.getIndex()), self());
 622
 623             } else {
 624                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
 625
 626                 // Local persistence is complete so send the CheckConsensusReached message to the behavior (which
 627                 // normally should still be the leader) to check if consensus has now been reached in conjunction with
 628                 // follower replication.
 629                 getCurrentBehavior().handleMessage(getSelf(), CheckConsensusReached.INSTANCE);
 630             }
 631         }, true);
 632
 633         if (wasAppended && hasFollowers()) {
 634             // Send log entry for replication.
 635             getCurrentBehavior().handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry,
 636                     !batchHint));
 637         }
 638     }
 639
 640     private ReplicatedLog replicatedLog() {
 641         return context.getReplicatedLog();
 642     }
 643
 644     protected String getId() {
 645         return context.getId();
 646     }
 647
 648     @VisibleForTesting
 649     void setCurrentBehavior(final RaftActorBehavior behavior) {
 650         context.setCurrentBehavior(behavior);
 651     }
 652
 653     protected RaftActorBehavior getCurrentBehavior() {
 654         return context.getCurrentBehavior();
 655     }
 656
 657     /**
 658      * Derived actors can call the isLeader method to check if the current
 659      * RaftActor is the Leader or not.
 660      *
 661      * @return true it this RaftActor is a Leader false otherwise
 662      */
 663     protected boolean isLeader() {
 664         return context.getId().equals(getCurrentBehavior().getLeaderId());
 665     }
 666
 667     protected final boolean isLeaderActive() {
 668         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
 669                 && !shuttingDown && !isLeadershipTransferInProgress();
 670     }
 671
 672     protected boolean isLeadershipTransferInProgress() {
 673         RaftActorLeadershipTransferCohort leadershipTransferInProgress = context.getRaftActorLeadershipTransferCohort();
 674         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 675     }
 676
 677     /**
 678      * Derived actor can call getLeader if they need a reference to the Leader.
 679      * This would be useful for example in forwarding a request to an actor
 680      * which is the leader
 681      *
 682      * @return A reference to the leader if known, null otherwise
 683      */
 684     public ActorSelection getLeader() {
 685         String leaderAddress = getLeaderAddress();
 686
 687         if (leaderAddress == null) {
 688             return null;
 689         }
 690
 691         return context.actorSelection(leaderAddress);
 692     }
 693
 694     /**
 695      * Returns the id of the current leader.
 696      *
 697      * @return the current leader's id
 698      */
 699     protected final String getLeaderId() {
 700         return getCurrentBehavior().getLeaderId();
 701     }
 702
 703     @VisibleForTesting
 704     protected final RaftState getRaftState() {
 705         return getCurrentBehavior().state();
 706     }
 707
 708     protected Long getCurrentTerm() {
 709         return context.getTermInformation().getCurrentTerm();
 710     }
 711
 712     protected RaftActorContext getRaftActorContext() {
 713         return context;
 714     }
 715
 716     protected void updateConfigParams(final ConfigParams configParams) {
 717
 718         // obtain the RaftPolicy for oldConfigParams and the updated one.
 719         String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
 720         String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
 721
 722         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 723             oldRaftPolicy, newRaftPolicy);
 724         context.setConfigParams(configParams);
 725         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 726             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 727             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 728             // avoids potential disruption. Otherwise, switch to Follower normally.
 729             RaftActorBehavior behavior = getCurrentBehavior();
 730             if (behavior != null && behavior.state() == RaftState.Follower) {
 731                 String previousLeaderId = behavior.getLeaderId();
 732                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 733
 734                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
 735                         previousLeaderId);
 736
 737                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 738             } else {
 739                 initializeBehavior();
 740             }
 741         }
 742     }
 743
 744     public final DataPersistenceProvider persistence() {
 745         return delegatingPersistenceProvider.getDelegate();
 746     }
 747
 748     public void setPersistence(final DataPersistenceProvider provider) {
 749         delegatingPersistenceProvider.setDelegate(provider);
 750     }
 751
 752     protected void setPersistence(final boolean persistent) {
 753         DataPersistenceProvider currentPersistence = persistence();
 754         if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
 755             setPersistence(new PersistentDataProvider(this));
 756
 757             if (getCurrentBehavior() != null) {
 758                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
 759                 captureSnapshot();
 760             }
 761         } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
 762             setPersistence(new NonPersistentDataProvider(this) {
 763                 /*
 764                  * The way snapshotting works is,
 765                  * <ol>
 766                  * <li> RaftActor calls createSnapshot on the Shard
 767                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 768                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 769                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 770                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 771                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 772                  * </ol>
 773                  */
 774                 @Override
 775                 public void saveSnapshot(final Object object) {
 776                     // Make saving Snapshot successful
 777                     // Committing the snapshot here would end up calling commit in the creating state which would
 778                     // be a state violation. That's why now we send a message to commit the snapshot.
 779                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 780                 }
 781             });
 782         }
 783     }
 784
 785     /**
 786      * setPeerAddress sets the address of a known peer at a later time.
 787      *
 788      * <p>
 789      * This is to account for situations where a we know that a peer
 790      * exists but we do not know an address up-front. This may also be used in
 791      * situations where a known peer starts off in a different location and we
 792      * need to change it's address
 793      *
 794      * <p>
 795      * Note that if the peerId does not match the list of peers passed to
 796      * this actor during construction an IllegalStateException will be thrown.
 797      */
 798     protected void setPeerAddress(final String peerId, final String peerAddress) {
 799         context.setPeerAddress(peerId, peerAddress);
 800     }
 801
 802     /**
 803      * The applyState method will be called by the RaftActor when some data
 804      * needs to be applied to the actor's state.
 805      *
 806      * @param clientActor A reference to the client who sent this message. This
 807      *                    is the same reference that was passed to persistData
 808      *                    by the derived actor. clientActor may be null when
 809      *                    the RaftActor is behaving as a follower or during
 810      *                    recovery.
 811      * @param identifier  The identifier of the persisted data. This is also
 812      *                    the same identifier that was passed to persistData by
 813      *                    the derived actor. identifier may be null when
 814      *                    the RaftActor is behaving as a follower or during
 815      *                    recovery
 816      * @param data        A piece of data that was persisted by the persistData call.
 817      *                    This should NEVER be null.
 818      */
 819     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 820
 821     /**
 822      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 823      */
 824     protected abstract @NonNull RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 825
 826     /**
 827      * This method is called when recovery is complete.
 828      */
 829     protected abstract void onRecoveryComplete();
 830
 831     /**
 832      * Returns the RaftActorSnapshotCohort to participate in snapshot captures.
 833      */
 834     protected abstract @NonNull RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 835
 836     /**
 837      * This method will be called by the RaftActor when the state of the
 838      * RaftActor changes. The derived actor can then use methods like
 839      * isLeader or getLeader to do something useful
 840      */
 841     protected abstract void onStateChanged();
 842
 843     /**
 844      * Notifier Actor for this RaftActor to notify when a role change happens.
 845      *
 846      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 847      */
 848     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 849
 850     /**
 851      * This method is called on the leader when a voting change operation completes.
 852      */
 853     protected void onVotingStateChangeComplete() {
 854     }
 855
 856     /**
 857      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 858      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 859      * work prior to performing the operation. On completion of any work, the run method must be called on the
 860      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 861      * this actor's thread dispatcher as as it modifies internal state.
 862      *
 863      * <p>
 864      * The default implementation immediately runs the operation.
 865      *
 866      * @param operation the operation to run
 867      */
 868     protected void pauseLeader(final Runnable operation) {
 869         operation.run();
 870     }
 871
 872     /**
 873      * This method is invoked when the actions hooked to the leader becoming paused failed to execute and the leader
 874      * should resume normal operations.
 875      *
 876      * <p>
 877      * Note this method can be invoked even before the operation supplied to {@link #pauseLeader(Runnable)} is invoked.
 878      */
 879     protected void unpauseLeader() {
 880
 881     }
 882
 883     protected void onLeaderChanged(final String oldLeader, final String newLeader) {
 884     }
 885
 886     private String getLeaderAddress() {
 887         if (isLeader()) {
 888             return getSelf().path().toString();
 889         }
 890         String leaderId = getLeaderId();
 891         if (leaderId == null) {
 892             return null;
 893         }
 894         String peerAddress = context.getPeerAddress(leaderId);
 895         LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
 896
 897         return peerAddress;
 898     }
 899
 900     protected boolean hasFollowers() {
 901         return getRaftActorContext().hasFollowers();
 902     }
 903
 904     private void captureSnapshot() {
 905         SnapshotManager snapshotManager = context.getSnapshotManager();
 906
 907         if (!snapshotManager.isCapturing()) {
 908             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 909             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 910                 replicatedLog().last(), idx);
 911
 912             snapshotManager.captureWithForcedTrim(replicatedLog().last(), idx);
 913         }
 914     }
 915
 916     /**
 917      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 918      * in which case we need to step down.
 919      */
 920     void becomeNonVoting() {
 921         if (isLeader()) {
 922             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 923                 @Override
 924                 public void onSuccess(final ActorRef raftActorRef) {
 925                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 926                     ensureFollowerState();
 927                 }
 928
 929                 @Override
 930                 public void onFailure(final ActorRef raftActorRef) {
 931                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 932                     ensureFollowerState();
 933                 }
 934
 935                 private void ensureFollowerState() {
 936                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 937                     // switch to Follower so ensure that.
 938                     if (getRaftState() != RaftState.Follower) {
 939                         initializeBehavior();
 940                     }
 941                 }
 942             }, null, RaftActorLeadershipTransferCohort.USE_DEFAULT_LEADER_TIMEOUT);
 943         }
 944     }
 945
 946     /**
 947      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 948      */
 949     private abstract static class BehaviorState implements Immutable {
 950         @Nullable abstract RaftActorBehavior getBehavior();
 951
 952         @Nullable abstract String getLastValidLeaderId();
 953
 954         @Nullable abstract String getLastLeaderId();
 955
 956         abstract short getLeaderPayloadVersion();
 957     }
 958
 959     /**
 960      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
 961      */
 962     private static final class SimpleBehaviorState extends BehaviorState {
 963         private final RaftActorBehavior behavior;
 964         private final String lastValidLeaderId;
 965         private final String lastLeaderId;
 966         private final short leaderPayloadVersion;
 967
 968         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
 969                 final RaftActorBehavior behavior) {
 970             this.lastValidLeaderId = lastValidLeaderId;
 971             this.lastLeaderId = lastLeaderId;
 972             this.behavior = requireNonNull(behavior);
 973             leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 974         }
 975
 976         @Override
 977         RaftActorBehavior getBehavior() {
 978             return behavior;
 979         }
 980
 981         @Override
 982         String getLastValidLeaderId() {
 983             return lastValidLeaderId;
 984         }
 985
 986         @Override
 987         short getLeaderPayloadVersion() {
 988             return leaderPayloadVersion;
 989         }
 990
 991         @Override
 992         String getLastLeaderId() {
 993             return lastLeaderId;
 994         }
 995     }
 996
 997     /**
 998      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 999      * An instance is created for each RaftActor. It has two functions:
1000      * - it keeps track of the last leader ID we have encountered since we have been created
1001      * - it creates state capture needed to transition from one behavior to the next
1002      */
1003     private static final class BehaviorStateTracker {
1004         /**
1005          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
1006          * allowed before we receive the first message, we know the leader ID to be null.
1007          */
1008         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
1009             @Override
1010             RaftActorBehavior getBehavior() {
1011                 return null;
1012             }
1013
1014             @Override
1015             String getLastValidLeaderId() {
1016                 return null;
1017             }
1018
1019             @Override
1020             short getLeaderPayloadVersion() {
1021                 return -1;
1022             }
1023
1024             @Override
1025             String getLastLeaderId() {
1026                 return null;
1027             }
1028         };
1029
1030         private String lastValidLeaderId;
1031         private String lastLeaderId;
1032
1033         BehaviorState capture(final RaftActorBehavior behavior) {
1034             if (behavior == null) {
1035                 verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
1036                 return NULL_BEHAVIOR_STATE;
1037             }
1038
1039             lastLeaderId = behavior.getLeaderId();
1040             if (lastLeaderId != null) {
1041                 lastValidLeaderId = lastLeaderId;
1042             }
1043
1044             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
1045         }
1046     }
1047 }