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