opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java

   1 /*
   2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
   3  * Copyright (c) 2015 Brocade Communications Systems, Inc. and others.  All rights reserved.
   4  *
   5  * This program and the accompanying materials are made available under the
   6  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   7  * and is available at http://www.eclipse.org/legal/epl-v10.html
   8  */
   9
  10 package org.opendaylight.controller.cluster.raft;
  11
  12 import akka.actor.ActorRef;
  13 import akka.actor.ActorSelection;
  14 import akka.actor.PoisonPill;
  15 import com.google.common.annotations.VisibleForTesting;
  16 import com.google.common.base.Optional;
  17 import com.google.common.base.Preconditions;
  18 import com.google.common.base.Verify;
  19 import com.google.common.collect.Lists;
  20 import java.util.Collection;
  21 import java.util.HashMap;
  22 import java.util.List;
  23 import java.util.Map;
  24 import java.util.Objects;
  25 import java.util.concurrent.TimeUnit;
  26 import javax.annotation.Nonnull;
  27 import javax.annotation.Nullable;
  28 import org.apache.commons.lang3.time.DurationFormatUtils;
  29 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  30 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
  31 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
  32 import org.opendaylight.controller.cluster.PersistentDataProvider;
  33 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
  34 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  35 import org.opendaylight.controller.cluster.notifications.RoleChanged;
  36 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  37 import org.opendaylight.controller.cluster.raft.base.messages.CheckConsensusReached;
  38 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
  39 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
  40 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  41 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
  42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
  43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  44 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  45 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  48 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
  49 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
  50 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
  51 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
  52 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
  53 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
  54 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
  55 import org.opendaylight.controller.cluster.raft.persisted.SimpleReplicatedLogEntry;
  56 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  57 import org.opendaylight.yangtools.concepts.Identifier;
  58 import org.opendaylight.yangtools.concepts.Immutable;
  59
  60 /**
  61  * RaftActor encapsulates a state machine that needs to be kept synchronized
  62  * in a cluster. It implements the RAFT algorithm as described in the paper
  63  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  64  * In Search of an Understandable Consensus Algorithm</a>
  65  *
  66  * <p>
  67  * RaftActor has 3 states and each state has a certain behavior associated
  68  * with it. A Raft actor can behave as,
  69  * <ul>
  70  * <li> A Leader </li>
  71  * <li> A Follower (or) </li>
  72  * <li> A Candidate </li>
  73  * </ul>
  74  *
  75  * <p>
  76  * A RaftActor MUST be a Leader in order to accept requests from clients to
  77  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  78  * a Leader it is also responsible for ensuring that all followers ultimately
  79  * have the same log and therefore the same state machine as itself.
  80  *
  81  * <p>
  82  * The current behavior of a RaftActor determines how election for leadership
  83  * is initiated and how peer RaftActors react to request for votes.
  84  *
  85  * <p>
  86  * Each RaftActor also needs to know the current election term. It uses this
  87  * information for a couple of things. One is to simply figure out who it
  88  * voted for in the last election. Another is to figure out if the message
  89  * it received to update it's state is stale.
  90  *
  91  * <p>
  92  * The RaftActor uses akka-persistence to store it's replicated log.
  93  * Furthermore through it's behaviors a Raft Actor determines
  94  * <ul>
  95  * <li> when a log entry should be persisted </li>
  96  * <li> when a log entry should be applied to the state machine (and) </li>
  97  * <li> when a snapshot should be saved </li>
  98  * </ul>
  99  */
 100 public abstract class RaftActor extends AbstractUntypedPersistentActor {
 101
 102     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
 103
 104     /**
 105      * This context should NOT be passed directly to any other actor it is
 106      * only to be consumed by the RaftActorBehaviors.
 107      */
 108     private final RaftActorContextImpl context;
 109
 110     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
 111
 112     private final PersistentDataProvider persistentProvider;
 113
 114     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
 115
 116     private RaftActorRecoverySupport raftRecovery;
 117
 118     private RaftActorSnapshotMessageSupport snapshotSupport;
 119
 120     private RaftActorServerConfigurationSupport serverConfigurationSupport;
 121
 122     private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
 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 (!possiblyHandleBehaviorMessage(message)) {
 268             handleNonRaftCommand(message);
 269         }
 270     }
 271
 272     private boolean possiblyHandleBehaviorMessage(final Object message) {
 273         final RaftActorBehavior currentBehavior = getCurrentBehavior();
 274         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
 275
 276         // A behavior indicates that it processed the change by returning a reference to the next behavior
 277         // to be used. A null return indicates it has not processed the message and we should be passing it to
 278         // the subclass for handling.
 279         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
 280         if (nextBehavior != null) {
 281             switchBehavior(state, nextBehavior);
 282             return true;
 283         }
 284
 285         return false;
 286     }
 287
 288     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
 289         LOG.debug("{}: Initiating leader transfer", persistenceId());
 290
 291         if (leadershipTransferInProgress == null) {
 292             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
 293             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
 294                 @Override
 295                 public void onSuccess(ActorRef raftActorRef) {
 296                     leadershipTransferInProgress = null;
 297                 }
 298
 299                 @Override
 300                 public void onFailure(ActorRef raftActorRef) {
 301                     leadershipTransferInProgress = null;
 302                 }
 303             });
 304
 305             leadershipTransferInProgress.addOnComplete(onComplete);
 306             leadershipTransferInProgress.init();
 307         } else {
 308             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
 309             leadershipTransferInProgress.addOnComplete(onComplete);
 310         }
 311     }
 312
 313     private void onShutDown() {
 314         LOG.debug("{}: onShutDown", persistenceId());
 315
 316         if (shuttingDown) {
 317             return;
 318         }
 319
 320         shuttingDown = true;
 321
 322         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 323         if (currentBehavior.state() != RaftState.Leader) {
 324             // For non-leaders shutdown is a no-op
 325             self().tell(PoisonPill.getInstance(), self());
 326             return;
 327         }
 328
 329         if (context.hasFollowers()) {
 330             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 331                 @Override
 332                 public void onSuccess(ActorRef raftActorRef) {
 333                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
 334                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 335                 }
 336
 337                 @Override
 338                 public void onFailure(ActorRef raftActorRef) {
 339                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
 340                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
 341                 }
 342             });
 343         } else {
 344             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
 345                 @Override
 346                 protected void doRun() {
 347                     self().tell(PoisonPill.getInstance(), self());
 348                 }
 349
 350                 @Override
 351                 protected void doCancel() {
 352                     self().tell(PoisonPill.getInstance(), self());
 353                 }
 354             });
 355         }
 356     }
 357
 358     private void onLeaderTransitioning(final LeaderTransitioning leaderTransitioning) {
 359         LOG.debug("{}: onLeaderTransitioning: {}", persistenceId(), leaderTransitioning);
 360         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 361         if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()
 362                 && leaderTransitioning.getLeaderId().equals(getCurrentBehavior().getLeaderId())) {
 363             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
 364                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
 365         }
 366     }
 367
 368     private void switchBehavior(SwitchBehavior message) {
 369         if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
 370             RaftState newState = message.getNewState();
 371             if ( newState == RaftState.Leader || newState == RaftState.Follower) {
 372                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
 373                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
 374                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
 375             } else {
 376                 LOG.warn("Switching to behavior : {} - not supported", newState);
 377             }
 378         }
 379     }
 380
 381     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
 382         setCurrentBehavior(nextBehavior);
 383         handleBehaviorChange(oldBehaviorState, nextBehavior);
 384     }
 385
 386     @VisibleForTesting
 387     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
 388         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
 389     }
 390
 391     private void onGetOnDemandRaftStats() {
 392         // Debugging message to retrieve raft stats.
 393
 394         Map<String, String> peerAddresses = new HashMap<>();
 395         Map<String, Boolean> peerVotingStates = new HashMap<>();
 396         for (PeerInfo info: context.getPeers()) {
 397             peerVotingStates.put(info.getId(), info.isVoting());
 398             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
 399         }
 400
 401         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
 402         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 403                 .commitIndex(context.getCommitIndex())
 404                 .currentTerm(context.getTermInformation().getCurrentTerm())
 405                 .inMemoryJournalDataSize(replicatedLog().dataSize())
 406                 .inMemoryJournalLogSize(replicatedLog().size())
 407                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
 408                 .lastApplied(context.getLastApplied())
 409                 .lastIndex(replicatedLog().lastIndex())
 410                 .lastTerm(replicatedLog().lastTerm())
 411                 .leader(getLeaderId())
 412                 .raftState(currentBehavior.state().toString())
 413                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 414                 .snapshotIndex(replicatedLog().getSnapshotIndex())
 415                 .snapshotTerm(replicatedLog().getSnapshotTerm())
 416                 .votedFor(context.getTermInformation().getVotedFor())
 417                 .isVoting(context.isVotingMember())
 418                 .peerAddresses(peerAddresses)
 419                 .peerVotingStates(peerVotingStates)
 420                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
 421
 422         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
 423         if (lastLogEntry != null) {
 424             builder.lastLogIndex(lastLogEntry.getIndex());
 425             builder.lastLogTerm(lastLogEntry.getTerm());
 426         }
 427
 428         if (getCurrentBehavior() instanceof AbstractLeader) {
 429             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
 430             Collection<String> followerIds = leader.getFollowerIds();
 431             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 432             for (String id: followerIds) {
 433                 final FollowerLogInformation info = leader.getFollower(id);
 434                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 435                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
 436                         context.getPeerInfo(info.getId()).isVoting()));
 437             }
 438
 439             builder.followerInfoList(followerInfoList);
 440         }
 441
 442         sender().tell(builder.build(), self());
 443
 444     }
 445
 446     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
 447         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 448
 449         if (oldBehavior != currentBehavior) {
 450             onStateChanged();
 451         }
 452
 453         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
 454         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
 455         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 456
 457         // it can happen that the state has not changed but the leader has changed.
 458         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 459         if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
 460                 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
 461             if (roleChangeNotifier.isPresent()) {
 462                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
 463                         currentBehavior.getLeaderPayloadVersion()), getSelf());
 464             }
 465
 466             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
 467
 468             if (leadershipTransferInProgress != null) {
 469                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
 470             }
 471
 472             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
 473         }
 474
 475         if (roleChangeNotifier.isPresent()
 476                 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
 477             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 478                     currentBehavior.state().name()), getSelf());
 479         }
 480     }
 481
 482     private void handleApplyState(ApplyState applyState) {
 483         long startTime = System.nanoTime();
 484
 485         Payload payload = applyState.getReplicatedLogEntry().getData();
 486         if (LOG.isDebugEnabled()) {
 487             LOG.debug("{}: Applying state for log index {} data {}",
 488                 persistenceId(), applyState.getReplicatedLogEntry().getIndex(), payload);
 489         }
 490
 491         if (!(payload instanceof NoopPayload) && !(payload instanceof ServerConfigurationPayload)) {
 492             applyState(applyState.getClientActor(), applyState.getIdentifier(), payload);
 493         }
 494
 495         long elapsedTime = System.nanoTime() - startTime;
 496         if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
 497             LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 498                     TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 499         }
 500
 501         // Send the ApplyState message back to self to handle further processing asynchronously.
 502         self().tell(applyState, self());
 503     }
 504
 505     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 506         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 507     }
 508
 509     @Override
 510     public long snapshotSequenceNr() {
 511         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
 512         // so that we can delete the persistent journal based on the saved sequence-number
 513         // However , when akka replays the journal during recovery, it replays it from the sequence number when the
 514         // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
 515         // last-sequence number known to us.
 516         return context.getSnapshotManager().getLastSequenceNumber();
 517     }
 518
 519     /**
 520      * Persists the given Payload in the journal and replicates to any followers. After successful completion,
 521      * {@link #applyState(ActorRef, Identifier, Object)} is notified.
 522      *
 523      * @param clientActor optional ActorRef that is provided via the applyState callback
 524      * @param identifier the payload identifier
 525      * @param data the payload data to persist
 526      * @param batchHint if true, an attempt is made to delay immediate replication and batch the payload with
 527      *        subsequent payloads for efficiency. Otherwise the payload is immediately replicated.
 528      */
 529     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data,
 530             final boolean batchHint) {
 531         ReplicatedLogEntry replicatedLogEntry = new SimpleReplicatedLogEntry(
 532             context.getReplicatedLog().lastIndex() + 1,
 533             context.getTermInformation().getCurrentTerm(), data);
 534         replicatedLogEntry.setPersistencePending(true);
 535
 536         LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 537
 538         final RaftActorContext raftContext = getRaftActorContext();
 539
 540         boolean wasAppended = replicatedLog().appendAndPersist(replicatedLogEntry, persistedLogEntry -> {
 541             // Clear the persistence pending flag in the log entry.
 542             persistedLogEntry.setPersistencePending(false);
 543
 544             if (!hasFollowers()) {
 545                 // Increment the Commit Index and the Last Applied values
 546                 raftContext.setCommitIndex(persistedLogEntry.getIndex());
 547                 raftContext.setLastApplied(persistedLogEntry.getIndex());
 548
 549                 // Apply the state immediately.
 550                 handleApplyState(new ApplyState(clientActor, identifier, persistedLogEntry));
 551
 552                 // Send a ApplyJournalEntries message so that we write the fact that we applied
 553                 // the state to durable storage
 554                 self().tell(new ApplyJournalEntries(persistedLogEntry.getIndex()), self());
 555
 556             } else {
 557                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
 558
 559                 // Local persistence is complete so send the CheckConsensusReached message to the behavior (which
 560                 // normally should still be the leader) to check if consensus has now been reached in conjunction with
 561                 // follower replication.
 562                 getCurrentBehavior().handleMessage(getSelf(), CheckConsensusReached.INSTANCE);
 563             }
 564         }, true);
 565
 566         if (wasAppended && hasFollowers()) {
 567             // Send log entry for replication.
 568             getCurrentBehavior().handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry,
 569                     !batchHint));
 570         }
 571     }
 572
 573     private ReplicatedLog replicatedLog() {
 574         return context.getReplicatedLog();
 575     }
 576
 577     protected String getId() {
 578         return context.getId();
 579     }
 580
 581     @VisibleForTesting
 582     void setCurrentBehavior(RaftActorBehavior behavior) {
 583         context.setCurrentBehavior(behavior);
 584     }
 585
 586     protected RaftActorBehavior getCurrentBehavior() {
 587         return context.getCurrentBehavior();
 588     }
 589
 590     /**
 591      * Derived actors can call the isLeader method to check if the current
 592      * RaftActor is the Leader or not.
 593      *
 594      * @return true it this RaftActor is a Leader false otherwise
 595      */
 596     protected boolean isLeader() {
 597         return context.getId().equals(getCurrentBehavior().getLeaderId());
 598     }
 599
 600     protected final boolean isLeaderActive() {
 601         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
 602                 && !shuttingDown && !isLeadershipTransferInProgress();
 603     }
 604
 605     private boolean isLeadershipTransferInProgress() {
 606         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
 607     }
 608
 609     /**
 610      * Derived actor can call getLeader if they need a reference to the Leader.
 611      * This would be useful for example in forwarding a request to an actor
 612      * which is the leader
 613      *
 614      * @return A reference to the leader if known, null otherwise
 615      */
 616     public ActorSelection getLeader() {
 617         String leaderAddress = getLeaderAddress();
 618
 619         if (leaderAddress == null) {
 620             return null;
 621         }
 622
 623         return context.actorSelection(leaderAddress);
 624     }
 625
 626     /**
 627      * Returns the id of the current leader.
 628      *
 629      * @return the current leader's id
 630      */
 631     protected final String getLeaderId() {
 632         return getCurrentBehavior().getLeaderId();
 633     }
 634
 635     @VisibleForTesting
 636     protected final RaftState getRaftState() {
 637         return getCurrentBehavior().state();
 638     }
 639
 640     protected Long getCurrentTerm() {
 641         return context.getTermInformation().getCurrentTerm();
 642     }
 643
 644     protected RaftActorContext getRaftActorContext() {
 645         return context;
 646     }
 647
 648     protected void updateConfigParams(ConfigParams configParams) {
 649
 650         // obtain the RaftPolicy for oldConfigParams and the updated one.
 651         String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
 652         String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
 653
 654         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
 655             oldRaftPolicy, newRaftPolicy);
 656         context.setConfigParams(configParams);
 657         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
 658             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
 659             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
 660             // avoids potential disruption. Otherwise, switch to Follower normally.
 661             RaftActorBehavior behavior = getCurrentBehavior();
 662             if (behavior != null && behavior.state() == RaftState.Follower) {
 663                 String previousLeaderId = behavior.getLeaderId();
 664                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
 665
 666                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
 667                         previousLeaderId);
 668
 669                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
 670             } else {
 671                 initializeBehavior();
 672             }
 673         }
 674     }
 675
 676     public final DataPersistenceProvider persistence() {
 677         return delegatingPersistenceProvider.getDelegate();
 678     }
 679
 680     public void setPersistence(DataPersistenceProvider provider) {
 681         delegatingPersistenceProvider.setDelegate(provider);
 682     }
 683
 684     protected void setPersistence(boolean persistent) {
 685         DataPersistenceProvider currentPersistence = persistence();
 686         if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
 687             setPersistence(new PersistentDataProvider(this));
 688
 689             if (getCurrentBehavior() != null) {
 690                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
 691                 captureSnapshot();
 692             }
 693         } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
 694             setPersistence(new NonPersistentDataProvider() {
 695                 /**
 696                  * The way snapshotting works is,
 697                  * <ol>
 698                  * <li> RaftActor calls createSnapshot on the Shard
 699                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 700                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
 701                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
 702                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
 703                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
 704                  * </ol>
 705                  */
 706                 @Override
 707                 public void saveSnapshot(Object object) {
 708                     // Make saving Snapshot successful
 709                     // Committing the snapshot here would end up calling commit in the creating state which would
 710                     // be a state violation. That's why now we send a message to commit the snapshot.
 711                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
 712                 }
 713             });
 714         }
 715     }
 716
 717     /**
 718      * setPeerAddress sets the address of a known peer at a later time.
 719      *
 720      * <p>
 721      * This is to account for situations where a we know that a peer
 722      * exists but we do not know an address up-front. This may also be used in
 723      * situations where a known peer starts off in a different location and we
 724      * need to change it's address
 725      *
 726      * <p>
 727      * Note that if the peerId does not match the list of peers passed to
 728      * this actor during construction an IllegalStateException will be thrown.
 729      */
 730     protected void setPeerAddress(String peerId, String peerAddress) {
 731         context.setPeerAddress(peerId, peerAddress);
 732     }
 733
 734     /**
 735      * The applyState method will be called by the RaftActor when some data
 736      * needs to be applied to the actor's state.
 737      *
 738      * @param clientActor A reference to the client who sent this message. This
 739      *                    is the same reference that was passed to persistData
 740      *                    by the derived actor. clientActor may be null when
 741      *                    the RaftActor is behaving as a follower or during
 742      *                    recovery.
 743      * @param identifier  The identifier of the persisted data. This is also
 744      *                    the same identifier that was passed to persistData by
 745      *                    the derived actor. identifier may be null when
 746      *                    the RaftActor is behaving as a follower or during
 747      *                    recovery
 748      * @param data        A piece of data that was persisted by the persistData call.
 749      *                    This should NEVER be null.
 750      */
 751     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
 752
 753     /**
 754      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
 755      */
 756     @Nonnull
 757     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
 758
 759     /**
 760      * This method is called when recovery is complete.
 761      */
 762     protected abstract void onRecoveryComplete();
 763
 764     /**
 765      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
 766      */
 767     @Nonnull
 768     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
 769
 770     /**
 771      * This method will be called by the RaftActor when the state of the
 772      * RaftActor changes. The derived actor can then use methods like
 773      * isLeader or getLeader to do something useful
 774      */
 775     protected abstract void onStateChanged();
 776
 777     /**
 778      * Notifier Actor for this RaftActor to notify when a role change happens.
 779      *
 780      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 781      */
 782     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 783
 784     /**
 785      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
 786      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
 787      * work prior to performing the operation. On completion of any work, the run method must be called on the
 788      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
 789      * this actor's thread dispatcher as as it modifies internal state.
 790      *
 791      * <p>
 792      * The default implementation immediately runs the operation.
 793      *
 794      * @param operation the operation to run
 795      */
 796     protected void pauseLeader(Runnable operation) {
 797         operation.run();
 798     }
 799
 800     protected void onLeaderChanged(String oldLeader, String newLeader) {
 801     }
 802
 803     private String getLeaderAddress() {
 804         if (isLeader()) {
 805             return getSelf().path().toString();
 806         }
 807         String leaderId = getLeaderId();
 808         if (leaderId == null) {
 809             return null;
 810         }
 811         String peerAddress = context.getPeerAddress(leaderId);
 812         LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
 813
 814         return peerAddress;
 815     }
 816
 817     protected boolean hasFollowers() {
 818         return getRaftActorContext().hasFollowers();
 819     }
 820
 821     private void captureSnapshot() {
 822         SnapshotManager snapshotManager = context.getSnapshotManager();
 823
 824         if (!snapshotManager.isCapturing()) {
 825             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
 826             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
 827                 replicatedLog().last(), idx);
 828
 829             snapshotManager.capture(replicatedLog().last(), idx);
 830         }
 831     }
 832
 833     /**
 834      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
 835      * in which case we need to step down.
 836      */
 837     void becomeNonVoting() {
 838         if (isLeader()) {
 839             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
 840                 @Override
 841                 public void onSuccess(ActorRef raftActorRef) {
 842                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
 843                     ensureFollowerState();
 844                 }
 845
 846                 @Override
 847                 public void onFailure(ActorRef raftActorRef) {
 848                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
 849                     ensureFollowerState();
 850                 }
 851
 852                 private void ensureFollowerState() {
 853                     // Whether or not leadership transfer succeeded, we have to step down as leader and
 854                     // switch to Follower so ensure that.
 855                     if (getRaftState() != RaftState.Follower) {
 856                         initializeBehavior();
 857                     }
 858                 }
 859             });
 860         }
 861     }
 862
 863     /**
 864      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
 865      */
 866     private abstract static class BehaviorState implements Immutable {
 867         @Nullable abstract RaftActorBehavior getBehavior();
 868
 869         @Nullable abstract String getLastValidLeaderId();
 870
 871         @Nullable abstract String getLastLeaderId();
 872
 873         @Nullable abstract short getLeaderPayloadVersion();
 874     }
 875
 876     /**
 877      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
 878      */
 879     private static final class SimpleBehaviorState extends BehaviorState {
 880         private final RaftActorBehavior behavior;
 881         private final String lastValidLeaderId;
 882         private final String lastLeaderId;
 883         private final short leaderPayloadVersion;
 884
 885         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
 886                 final RaftActorBehavior behavior) {
 887             this.lastValidLeaderId = lastValidLeaderId;
 888             this.lastLeaderId = lastLeaderId;
 889             this.behavior = Preconditions.checkNotNull(behavior);
 890             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
 891         }
 892
 893         @Override
 894         RaftActorBehavior getBehavior() {
 895             return behavior;
 896         }
 897
 898         @Override
 899         String getLastValidLeaderId() {
 900             return lastValidLeaderId;
 901         }
 902
 903         @Override
 904         short getLeaderPayloadVersion() {
 905             return leaderPayloadVersion;
 906         }
 907
 908         @Override
 909         String getLastLeaderId() {
 910             return lastLeaderId;
 911         }
 912     }
 913
 914     /**
 915      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
 916      * An instance is created for each RaftActor. It has two functions:
 917      * - it keeps track of the last leader ID we have encountered since we have been created
 918      * - it creates state capture needed to transition from one behavior to the next
 919      */
 920     private static final class BehaviorStateTracker {
 921         /**
 922          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
 923          * allowed before we receive the first message, we know the leader ID to be null.
 924          */
 925         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
 926             @Override
 927             RaftActorBehavior getBehavior() {
 928                 return null;
 929             }
 930
 931             @Override
 932             String getLastValidLeaderId() {
 933                 return null;
 934             }
 935
 936             @Override
 937             short getLeaderPayloadVersion() {
 938                 return -1;
 939             }
 940
 941             @Override
 942             String getLastLeaderId() {
 943                 return null;
 944             }
 945         };
 946
 947         private String lastValidLeaderId;
 948         private String lastLeaderId;
 949
 950         BehaviorState capture(final RaftActorBehavior behavior) {
 951             if (behavior == null) {
 952                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
 953                 return NULL_BEHAVIOR_STATE;
 954             }
 955
 956             lastLeaderId = behavior.getLeaderId();
 957             if (lastLeaderId != null) {
 958                 lastValidLeaderId = lastLeaderId;
 959             }
 960
 961             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
 962         }
 963     }
 964
 965 }