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