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  *
   4  * This program and the accompanying materials are made available under the
   5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   6  * and is available at http://www.eclipse.org/legal/epl-v10.html
   7  */
   8
   9 package org.opendaylight.controller.cluster.raft;
  10
  11 import akka.actor.ActorRef;
  12 import akka.actor.ActorSelection;
  13 import akka.japi.Procedure;
  14 import akka.persistence.RecoveryCompleted;
  15 import akka.persistence.SaveSnapshotFailure;
  16 import akka.persistence.SaveSnapshotSuccess;
  17 import akka.persistence.SnapshotOffer;
  18 import akka.persistence.SnapshotSelectionCriteria;
  19 import com.google.common.annotations.VisibleForTesting;
  20 import com.google.common.base.Objects;
  21 import com.google.common.base.Optional;
  22 import com.google.common.base.Stopwatch;
  23 import com.google.common.collect.ImmutableMap;
  24 import com.google.common.collect.Lists;
  25 import com.google.protobuf.ByteString;
  26 import java.io.Serializable;
  27 import java.util.Collection;
  28 import java.util.List;
  29 import java.util.Map;
  30 import java.util.concurrent.TimeUnit;
  31 import org.apache.commons.lang3.time.DurationFormatUtils;
  32 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  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.ApplyJournalEntries;
  37 import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
  38 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
  39 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  40 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
  41 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
  42 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  43 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  44 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
  45 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  46 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  47 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  48 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  49 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  50 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
  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.protobuff.client.messages.Payload;
  54 import org.slf4j.Logger;
  55 import org.slf4j.LoggerFactory;
  56
  57 /**
  58  * RaftActor encapsulates a state machine that needs to be kept synchronized
  59  * in a cluster. It implements the RAFT algorithm as described in the paper
  60  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  61  * In Search of an Understandable Consensus Algorithm</a>
  62  * <p/>
  63  * RaftActor has 3 states and each state has a certain behavior associated
  64  * with it. A Raft actor can behave as,
  65  * <ul>
  66  * <li> A Leader </li>
  67  * <li> A Follower (or) </li>
  68  * <li> A Candidate </li>
  69  * </ul>
  70  * <p/>
  71  * <p/>
  72  * A RaftActor MUST be a Leader in order to accept requests from clients to
  73  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  74  * a Leader it is also responsible for ensuring that all followers ultimately
  75  * have the same log and therefore the same state machine as itself.
  76  * <p/>
  77  * <p/>
  78  * The current behavior of a RaftActor determines how election for leadership
  79  * is initiated and how peer RaftActors react to request for votes.
  80  * <p/>
  81  * <p/>
  82  * Each RaftActor also needs to know the current election term. It uses this
  83  * information for a couple of things. One is to simply figure out who it
  84  * voted for in the last election. Another is to figure out if the message
  85  * it received to update it's state is stale.
  86  * <p/>
  87  * <p/>
  88  * The RaftActor uses akka-persistence to store it's replicated log.
  89  * Furthermore through it's behaviors a Raft Actor determines
  90  * <p/>
  91  * <ul>
  92  * <li> when a log entry should be persisted </li>
  93  * <li> when a log entry should be applied to the state machine (and) </li>
  94  * <li> when a snapshot should be saved </li>
  95  * </ul>
  96  */
  97 public abstract class RaftActor extends AbstractUntypedPersistentActor {
  98
  99     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
 100
 101     private static final Procedure<ApplyJournalEntries> APPLY_JOURNAL_ENTRIES_PERSIST_CALLBACK =
 102             new Procedure<ApplyJournalEntries>() {
 103                 @Override
 104                 public void apply(ApplyJournalEntries param) throws Exception {
 105                 }
 106             };
 107
 108     protected final Logger LOG = LoggerFactory.getLogger(getClass());
 109
 110     /**
 111      * The current state determines the current behavior of a RaftActor
 112      * A Raft Actor always starts off in the Follower State
 113      */
 114     private RaftActorBehavior currentBehavior;
 115
 116     /**
 117      * This context should NOT be passed directly to any other actor it is
 118      * only to be consumed by the RaftActorBehaviors
 119      */
 120     private final RaftActorContextImpl context;
 121
 122     /**
 123      * The in-memory journal
 124      */
 125     private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
 126
 127     private CaptureSnapshot captureSnapshot = null;
 128
 129     private Stopwatch recoveryTimer;
 130
 131     private int currentRecoveryBatchCount;
 132
 133     private final BehaviorStateHolder reusableBehaviorStateHolder = new BehaviorStateHolder();
 134
 135     public RaftActor(String id, Map<String, String> peerAddresses) {
 136         this(id, peerAddresses, Optional.<ConfigParams>absent());
 137     }
 138
 139     public RaftActor(String id, Map<String, String> peerAddresses,
 140          Optional<ConfigParams> configParams) {
 141
 142         context = new RaftActorContextImpl(this.getSelf(),
 143             this.getContext(), id, new ElectionTermImpl(),
 144             -1, -1, replicatedLog, peerAddresses,
 145             (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
 146             LOG);
 147     }
 148
 149     private void initRecoveryTimer() {
 150         if(recoveryTimer == null) {
 151             recoveryTimer = Stopwatch.createStarted();
 152         }
 153     }
 154
 155     @Override
 156     public void preStart() throws Exception {
 157         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 158                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 159
 160         super.preStart();
 161     }
 162
 163     @Override
 164     public void postStop() {
 165         if(currentBehavior != null) {
 166             try {
 167                 currentBehavior.close();
 168             } catch (Exception e) {
 169                 LOG.debug("{}: Error closing behavior {}", persistenceId(), currentBehavior.state());
 170             }
 171         }
 172
 173         super.postStop();
 174     }
 175
 176     @Override
 177     public void handleRecover(Object message) {
 178         if(persistence().isRecoveryApplicable()) {
 179             if (message instanceof SnapshotOffer) {
 180                 onRecoveredSnapshot((SnapshotOffer) message);
 181             } else if (message instanceof ReplicatedLogEntry) {
 182                 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
 183             } else if (message instanceof ApplyLogEntries) {
 184                 // Handle this message for backwards compatibility with pre-Lithium versions.
 185                 onRecoveredApplyLogEntries(((ApplyLogEntries) message).getToIndex());
 186             } else if (message instanceof ApplyJournalEntries) {
 187                 onRecoveredApplyLogEntries(((ApplyJournalEntries) message).getToIndex());
 188             } else if (message instanceof DeleteEntries) {
 189                 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
 190             } else if (message instanceof UpdateElectionTerm) {
 191                 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
 192                         ((UpdateElectionTerm) message).getVotedFor());
 193             } else if (message instanceof RecoveryCompleted) {
 194                 onRecoveryCompletedMessage();
 195             }
 196         } else {
 197             if (message instanceof RecoveryCompleted) {
 198                 // Delete all the messages from the akka journal so that we do not end up with consistency issues
 199                 // Note I am not using the dataPersistenceProvider and directly using the akka api here
 200                 deleteMessages(lastSequenceNr());
 201
 202                 // Delete all the akka snapshots as they will not be needed
 203                 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
 204
 205                 onRecoveryComplete();
 206
 207                 initializeBehavior();
 208             }
 209         }
 210     }
 211
 212     private void onRecoveredSnapshot(SnapshotOffer offer) {
 213         if(LOG.isDebugEnabled()) {
 214             LOG.debug("{}: SnapshotOffer called..", persistenceId());
 215         }
 216
 217         initRecoveryTimer();
 218
 219         Snapshot snapshot = (Snapshot) offer.snapshot();
 220
 221         // Create a replicated log with the snapshot information
 222         // The replicated log can be used later on to retrieve this snapshot
 223         // when we need to install it on a peer
 224         replicatedLog = new ReplicatedLogImpl(snapshot);
 225
 226         context.setReplicatedLog(replicatedLog);
 227         context.setLastApplied(snapshot.getLastAppliedIndex());
 228         context.setCommitIndex(snapshot.getLastAppliedIndex());
 229
 230         Stopwatch timer = Stopwatch.createStarted();
 231
 232         // Apply the snapshot to the actors state
 233         applyRecoverySnapshot(snapshot.getState());
 234
 235         timer.stop();
 236         LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
 237                 replicatedLog.size(), persistenceId(), timer.toString(),
 238                 replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm());
 239     }
 240
 241     private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
 242         if(LOG.isDebugEnabled()) {
 243             LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex());
 244         }
 245
 246         replicatedLog.append(logEntry);
 247     }
 248
 249     private void onRecoveredApplyLogEntries(long toIndex) {
 250         if(LOG.isDebugEnabled()) {
 251             LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}",
 252                     persistenceId(), context.getLastApplied() + 1, toIndex);
 253         }
 254
 255         for (long i = context.getLastApplied() + 1; i <= toIndex; i++) {
 256             batchRecoveredLogEntry(replicatedLog.get(i));
 257         }
 258
 259         context.setLastApplied(toIndex);
 260         context.setCommitIndex(toIndex);
 261     }
 262
 263     private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
 264         initRecoveryTimer();
 265
 266         int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
 267         if(currentRecoveryBatchCount == 0) {
 268             startLogRecoveryBatch(batchSize);
 269         }
 270
 271         appendRecoveredLogEntry(logEntry.getData());
 272
 273         if(++currentRecoveryBatchCount >= batchSize) {
 274             endCurrentLogRecoveryBatch();
 275         }
 276     }
 277
 278     private void endCurrentLogRecoveryBatch() {
 279         applyCurrentLogRecoveryBatch();
 280         currentRecoveryBatchCount = 0;
 281     }
 282
 283     private void onRecoveryCompletedMessage() {
 284         if(currentRecoveryBatchCount > 0) {
 285             endCurrentLogRecoveryBatch();
 286         }
 287
 288         onRecoveryComplete();
 289
 290         String recoveryTime = "";
 291         if(recoveryTimer != null) {
 292             recoveryTimer.stop();
 293             recoveryTime = " in " + recoveryTimer.toString();
 294             recoveryTimer = null;
 295         }
 296
 297         LOG.info(
 298             "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
 299                 "Persistence Id =  " + persistenceId() +
 300                 " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
 301                 "journal-size={}",
 302             replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(),
 303             replicatedLog.getSnapshotTerm(), replicatedLog.size());
 304
 305         initializeBehavior();
 306     }
 307
 308     protected void initializeBehavior(){
 309         changeCurrentBehavior(new Follower(context));
 310     }
 311
 312     protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
 313         reusableBehaviorStateHolder.init(currentBehavior);
 314         currentBehavior = newBehavior;
 315         handleBehaviorChange(reusableBehaviorStateHolder, currentBehavior);
 316     }
 317
 318     @Override public void handleCommand(Object message) {
 319         if (message instanceof ApplyState){
 320             ApplyState applyState = (ApplyState) message;
 321
 322             long elapsedTime = (System.nanoTime() - applyState.getStartTime());
 323             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 324                 LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 325                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 326             }
 327
 328             if(LOG.isDebugEnabled()) {
 329                 LOG.debug("{}: Applying state for log index {} data {}",
 330                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 331                     applyState.getReplicatedLogEntry().getData());
 332             }
 333
 334             applyState(applyState.getClientActor(), applyState.getIdentifier(),
 335                 applyState.getReplicatedLogEntry().getData());
 336
 337         } else if (message instanceof ApplyJournalEntries){
 338             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 339             if(LOG.isDebugEnabled()) {
 340                 LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), applyEntries.getToIndex());
 341             }
 342
 343             persistence().persist(applyEntries, APPLY_JOURNAL_ENTRIES_PERSIST_CALLBACK);
 344
 345         } else if(message instanceof ApplySnapshot ) {
 346             Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
 347
 348             if(LOG.isDebugEnabled()) {
 349                 LOG.debug("{}: ApplySnapshot called on Follower Actor " +
 350                         "snapshotIndex:{}, snapshotTerm:{}", persistenceId(), snapshot.getLastAppliedIndex(),
 351                     snapshot.getLastAppliedTerm()
 352                 );
 353             }
 354
 355             applySnapshot(snapshot.getState());
 356
 357             //clears the followers log, sets the snapshot index to ensure adjusted-index works
 358             replicatedLog = new ReplicatedLogImpl(snapshot);
 359             context.setReplicatedLog(replicatedLog);
 360             context.setLastApplied(snapshot.getLastAppliedIndex());
 361
 362         } else if (message instanceof FindLeader) {
 363             getSender().tell(
 364                 new FindLeaderReply(getLeaderAddress()),
 365                 getSelf()
 366             );
 367
 368         } else if (message instanceof SaveSnapshotSuccess) {
 369             SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
 370             LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId());
 371
 372             long sequenceNumber = success.metadata().sequenceNr();
 373
 374             commitSnapshot(sequenceNumber);
 375
 376         } else if (message instanceof SaveSnapshotFailure) {
 377             SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
 378
 379             LOG.error("{}: SaveSnapshotFailure received for snapshot Cause:",
 380                     persistenceId(), saveSnapshotFailure.cause());
 381
 382             context.getReplicatedLog().snapshotRollback();
 383
 384             LOG.info("{}: Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
 385                 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}", persistenceId(),
 386                 context.getReplicatedLog().getSnapshotIndex(),
 387                 context.getReplicatedLog().getSnapshotTerm(),
 388                 context.getReplicatedLog().size());
 389
 390         } else if (message instanceof CaptureSnapshot) {
 391             LOG.debug("{}: CaptureSnapshot received by actor: {}", persistenceId(), message);
 392
 393             if(captureSnapshot == null) {
 394                 captureSnapshot = (CaptureSnapshot)message;
 395                 createSnapshot();
 396             }
 397
 398         } else if (message instanceof CaptureSnapshotReply){
 399             handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot());
 400         } else if(message instanceof GetOnDemandRaftState) {
 401             onGetOnDemandRaftStats();
 402         } else {
 403             reusableBehaviorStateHolder.init(currentBehavior);
 404
 405             currentBehavior = currentBehavior.handleMessage(getSender(), message);
 406
 407             handleBehaviorChange(reusableBehaviorStateHolder, currentBehavior);
 408         }
 409     }
 410
 411     private void onGetOnDemandRaftStats() {
 412         // Debugging message to retrieve raft stats.
 413
 414         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 415                 .commitIndex(context.getCommitIndex())
 416                 .currentTerm(context.getTermInformation().getCurrentTerm())
 417                 .inMemoryJournalDataSize(replicatedLog.dataSize())
 418                 .inMemoryJournalLogSize(replicatedLog.size())
 419                 .isSnapshotCaptureInitiated(context.isSnapshotCaptureInitiated())
 420                 .lastApplied(context.getLastApplied())
 421                 .lastIndex(replicatedLog.lastIndex())
 422                 .lastTerm(replicatedLog.lastTerm())
 423                 .leader(getLeaderId())
 424                 .raftState(currentBehavior.state().toString())
 425                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 426                 .snapshotIndex(replicatedLog.getSnapshotIndex())
 427                 .snapshotTerm(replicatedLog.getSnapshotTerm())
 428                 .votedFor(context.getTermInformation().getVotedFor())
 429                 .peerAddresses(ImmutableMap.copyOf(context.getPeerAddresses()));
 430
 431         ReplicatedLogEntry lastLogEntry = getLastLogEntry();
 432         if (lastLogEntry != null) {
 433             builder.lastLogIndex(lastLogEntry.getIndex());
 434             builder.lastLogTerm(lastLogEntry.getTerm());
 435         }
 436
 437         if(currentBehavior instanceof AbstractLeader) {
 438             AbstractLeader leader = (AbstractLeader)currentBehavior;
 439             Collection<String> followerIds = leader.getFollowerIds();
 440             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 441             for(String id: followerIds) {
 442                 final FollowerLogInformation info = leader.getFollower(id);
 443                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 444                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity())));
 445             }
 446
 447             builder.followerInfoList(followerInfoList);
 448         }
 449
 450         sender().tell(builder.build(), self());
 451
 452     }
 453
 454     private void handleBehaviorChange(BehaviorStateHolder oldBehaviorState, RaftActorBehavior currentBehavior) {
 455         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
 456
 457         if (oldBehavior != currentBehavior){
 458             onStateChanged();
 459         }
 460
 461         String oldBehaviorLeaderId = oldBehavior == null ? null : oldBehaviorState.getLeaderId();
 462         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
 463
 464         // it can happen that the state has not changed but the leader has changed.
 465         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
 466         if(!Objects.equal(oldBehaviorLeaderId, currentBehavior.getLeaderId())) {
 467             if(roleChangeNotifier.isPresent()) {
 468                 roleChangeNotifier.get().tell(new LeaderStateChanged(getId(), currentBehavior.getLeaderId()), getSelf());
 469             }
 470
 471             onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
 472         }
 473
 474         if (roleChangeNotifier.isPresent() &&
 475                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
 476             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
 477                     currentBehavior.state().name()), getSelf());
 478         }
 479     }
 480
 481     /**
 482      * When a derived RaftActor needs to persist something it must call
 483      * persistData.
 484      *
 485      * @param clientActor
 486      * @param identifier
 487      * @param data
 488      */
 489     protected void persistData(final ActorRef clientActor, final String identifier,
 490         final Payload data) {
 491
 492         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 493             context.getReplicatedLog().lastIndex() + 1,
 494             context.getTermInformation().getCurrentTerm(), data);
 495
 496         if(LOG.isDebugEnabled()) {
 497             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 498         }
 499
 500         final RaftActorContext raftContext = getRaftActorContext();
 501
 502         replicatedLog
 503                 .appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
 504                     @Override
 505                     public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
 506                         if(!hasFollowers()){
 507                             // Increment the Commit Index and the Last Applied values
 508                             raftContext.setCommitIndex(replicatedLogEntry.getIndex());
 509                             raftContext.setLastApplied(replicatedLogEntry.getIndex());
 510
 511                             // Apply the state immediately
 512                             applyState(clientActor, identifier, data);
 513
 514                             // Send a ApplyJournalEntries message so that we write the fact that we applied
 515                             // the state to durable storage
 516                             self().tell(new ApplyJournalEntries(replicatedLogEntry.getIndex()), self());
 517
 518                             // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot
 519                             if(!context.isSnapshotCaptureInitiated()){
 520                                 raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(),
 521                                         raftContext.getTermInformation().getCurrentTerm());
 522                                 raftContext.getReplicatedLog().snapshotCommit();
 523                             } else {
 524                                 LOG.debug("{}: Skipping fake snapshotting for {} because real snapshotting is in progress",
 525                                         persistenceId(), getId());
 526                             }
 527                         } else if (clientActor != null) {
 528                             // Send message for replication
 529                             currentBehavior.handleMessage(getSelf(),
 530                                     new Replicate(clientActor, identifier,
 531                                             replicatedLogEntry)
 532                             );
 533                         }
 534
 535                     }
 536                 });    }
 537
 538     protected String getId() {
 539         return context.getId();
 540     }
 541
 542     /**
 543      * Derived actors can call the isLeader method to check if the current
 544      * RaftActor is the Leader or not
 545      *
 546      * @return true it this RaftActor is a Leader false otherwise
 547      */
 548     protected boolean isLeader() {
 549         return context.getId().equals(currentBehavior.getLeaderId());
 550     }
 551
 552     /**
 553      * Derived actor can call getLeader if they need a reference to the Leader.
 554      * This would be useful for example in forwarding a request to an actor
 555      * which is the leader
 556      *
 557      * @return A reference to the leader if known, null otherwise
 558      */
 559     protected ActorSelection getLeader(){
 560         String leaderAddress = getLeaderAddress();
 561
 562         if(leaderAddress == null){
 563             return null;
 564         }
 565
 566         return context.actorSelection(leaderAddress);
 567     }
 568
 569     /**
 570      *
 571      * @return the current leader's id
 572      */
 573     protected String getLeaderId(){
 574         return currentBehavior.getLeaderId();
 575     }
 576
 577     protected RaftState getRaftState() {
 578         return currentBehavior.state();
 579     }
 580
 581     protected ReplicatedLogEntry getLastLogEntry() {
 582         return replicatedLog.last();
 583     }
 584
 585     protected Long getCurrentTerm(){
 586         return context.getTermInformation().getCurrentTerm();
 587     }
 588
 589     protected Long getCommitIndex(){
 590         return context.getCommitIndex();
 591     }
 592
 593     protected Long getLastApplied(){
 594         return context.getLastApplied();
 595     }
 596
 597     protected RaftActorContext getRaftActorContext() {
 598         return context;
 599     }
 600
 601     protected void updateConfigParams(ConfigParams configParams) {
 602         context.setConfigParams(configParams);
 603     }
 604
 605     /**
 606      * setPeerAddress sets the address of a known peer at a later time.
 607      * <p>
 608      * This is to account for situations where a we know that a peer
 609      * exists but we do not know an address up-front. This may also be used in
 610      * situations where a known peer starts off in a different location and we
 611      * need to change it's address
 612      * <p>
 613      * Note that if the peerId does not match the list of peers passed to
 614      * this actor during construction an IllegalStateException will be thrown.
 615      *
 616      * @param peerId
 617      * @param peerAddress
 618      */
 619     protected void setPeerAddress(String peerId, String peerAddress){
 620         context.setPeerAddress(peerId, peerAddress);
 621     }
 622
 623     protected void commitSnapshot(long sequenceNumber) {
 624         context.getReplicatedLog().snapshotCommit();
 625
 626         // TODO: Not sure if we want to be this aggressive with trimming stuff
 627         trimPersistentData(sequenceNumber);
 628     }
 629
 630     /**
 631      * The applyState method will be called by the RaftActor when some data
 632      * needs to be applied to the actor's state
 633      *
 634      * @param clientActor A reference to the client who sent this message. This
 635      *                    is the same reference that was passed to persistData
 636      *                    by the derived actor. clientActor may be null when
 637      *                    the RaftActor is behaving as a follower or during
 638      *                    recovery.
 639      * @param identifier  The identifier of the persisted data. This is also
 640      *                    the same identifier that was passed to persistData by
 641      *                    the derived actor. identifier may be null when
 642      *                    the RaftActor is behaving as a follower or during
 643      *                    recovery
 644      * @param data        A piece of data that was persisted by the persistData call.
 645      *                    This should NEVER be null.
 646      */
 647     protected abstract void applyState(ActorRef clientActor, String identifier,
 648         Object data);
 649
 650     /**
 651      * This method is called during recovery at the start of a batch of state entries. Derived
 652      * classes should perform any initialization needed to start a batch.
 653      */
 654     protected abstract void startLogRecoveryBatch(int maxBatchSize);
 655
 656     /**
 657      * This method is called during recovery to append state data to the current batch. This method
 658      * is called 1 or more times after {@link #startLogRecoveryBatch}.
 659      *
 660      * @param data the state data
 661      */
 662     protected abstract void appendRecoveredLogEntry(Payload data);
 663
 664     /**
 665      * This method is called during recovery to reconstruct the state of the actor.
 666      *
 667      * @param snapshotBytes A snapshot of the state of the actor
 668      */
 669     protected abstract void applyRecoverySnapshot(byte[] snapshotBytes);
 670
 671     /**
 672      * This method is called during recovery at the end of a batch to apply the current batched
 673      * log entries. This method is called after {@link #appendRecoveredLogEntry}.
 674      */
 675     protected abstract void applyCurrentLogRecoveryBatch();
 676
 677     /**
 678      * This method is called when recovery is complete.
 679      */
 680     protected abstract void onRecoveryComplete();
 681
 682     /**
 683      * This method will be called by the RaftActor when a snapshot needs to be
 684      * created. The derived actor should respond with its current state.
 685      * <p/>
 686      * During recovery the state that is returned by the derived actor will
 687      * be passed back to it by calling the applySnapshot  method
 688      *
 689      * @return The current state of the actor
 690      */
 691     protected abstract void createSnapshot();
 692
 693     /**
 694      * This method can be called at any other point during normal
 695      * operations when the derived actor is out of sync with it's peers
 696      * and the only way to bring it in sync is by applying a snapshot
 697      *
 698      * @param snapshotBytes A snapshot of the state of the actor
 699      */
 700     protected abstract void applySnapshot(byte[] snapshotBytes);
 701
 702     /**
 703      * This method will be called by the RaftActor when the state of the
 704      * RaftActor changes. The derived actor can then use methods like
 705      * isLeader or getLeader to do something useful
 706      */
 707     protected abstract void onStateChanged();
 708
 709     protected abstract DataPersistenceProvider persistence();
 710
 711     /**
 712      * Notifier Actor for this RaftActor to notify when a role change happens
 713      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 714      */
 715     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 716
 717     protected void onLeaderChanged(String oldLeader, String newLeader){};
 718
 719     private void trimPersistentData(long sequenceNumber) {
 720         // Trim akka snapshots
 721         // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
 722         // For now guessing that it is ANDed.
 723         persistence().deleteSnapshots(new SnapshotSelectionCriteria(
 724             sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
 725
 726         // Trim akka journal
 727         persistence().deleteMessages(sequenceNumber);
 728     }
 729
 730     private String getLeaderAddress(){
 731         if(isLeader()){
 732             return getSelf().path().toString();
 733         }
 734         String leaderId = currentBehavior.getLeaderId();
 735         if (leaderId == null) {
 736             return null;
 737         }
 738         String peerAddress = context.getPeerAddress(leaderId);
 739         if(LOG.isDebugEnabled()) {
 740             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 741                     persistenceId(), leaderId, peerAddress);
 742         }
 743
 744         return peerAddress;
 745     }
 746
 747     private void handleCaptureSnapshotReply(byte[] snapshotBytes) {
 748         LOG.debug("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length);
 749
 750         // create a snapshot object from the state provided and save it
 751         // when snapshot is saved async, SaveSnapshotSuccess is raised.
 752
 753         Snapshot sn = Snapshot.create(snapshotBytes,
 754             context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
 755             captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
 756             captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
 757
 758         persistence().saveSnapshot(sn);
 759
 760         LOG.info("{}: Persisting of snapshot done:{}", persistenceId(), sn.getLogMessage());
 761
 762         long dataThreshold = getTotalMemory() *
 763                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 764         if (context.getReplicatedLog().dataSize() > dataThreshold) {
 765
 766             if(LOG.isDebugEnabled()) {
 767                 LOG.debug("{}: dataSize {} exceeds dataThreshold {} - doing snapshotPreCommit with index {}",
 768                         persistenceId(), context.getReplicatedLog().dataSize(), dataThreshold,
 769                         captureSnapshot.getLastAppliedIndex());
 770             }
 771
 772             // if memory is less, clear the log based on lastApplied.
 773             // this could/should only happen if one of the followers is down
 774             // as normally we keep removing from the log when its replicated to all.
 775             context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getLastAppliedIndex(),
 776                     captureSnapshot.getLastAppliedTerm());
 777
 778             // Don't reset replicatedToAllIndex to -1 as this may prevent us from trimming the log after an
 779             // install snapshot to a follower.
 780             if(captureSnapshot.getReplicatedToAllIndex() >= 0) {
 781                 getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
 782             }
 783         } else if(captureSnapshot.getReplicatedToAllIndex() != -1){
 784             // clear the log based on replicatedToAllIndex
 785             context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getReplicatedToAllIndex(),
 786                     captureSnapshot.getReplicatedToAllTerm());
 787
 788             getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
 789         } else {
 790             // The replicatedToAllIndex was not found in the log
 791             // This means that replicatedToAllIndex never moved beyond -1 or that it is already in the snapshot.
 792             // In this scenario we may need to save the snapshot to the akka persistence
 793             // snapshot for recovery but we do not need to do the replicated log trimming.
 794             context.getReplicatedLog().snapshotPreCommit(replicatedLog.getSnapshotIndex(),
 795                     replicatedLog.getSnapshotTerm());
 796         }
 797
 798
 799         LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex: {} " +
 800             "and term: {}", persistenceId(), replicatedLog.getSnapshotIndex(),
 801             replicatedLog.getSnapshotTerm());
 802
 803         if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
 804             // this would be call straight to the leader and won't initiate in serialization
 805             currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(
 806                     ByteString.copyFrom(snapshotBytes)));
 807         }
 808
 809         captureSnapshot = null;
 810         context.setSnapshotCaptureInitiated(false);
 811     }
 812
 813     protected long getTotalMemory() {
 814         return Runtime.getRuntime().totalMemory();
 815     }
 816
 817     protected boolean hasFollowers(){
 818         return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
 819     }
 820
 821     private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
 822
 823         private static final int DATA_SIZE_DIVIDER = 5;
 824         private long dataSizeSinceLastSnapshot = 0;
 825
 826         public ReplicatedLogImpl(Snapshot snapshot) {
 827             super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
 828                 snapshot.getUnAppliedEntries());
 829         }
 830
 831         public ReplicatedLogImpl() {
 832             super();
 833         }
 834
 835         @Override public void removeFromAndPersist(long logEntryIndex) {
 836             int adjustedIndex = adjustedIndex(logEntryIndex);
 837
 838             if (adjustedIndex < 0) {
 839                 return;
 840             }
 841
 842             // FIXME: Maybe this should be done after the command is saved
 843             journal.subList(adjustedIndex , journal.size()).clear();
 844
 845             persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>() {
 846
 847                 @Override
 848                 public void apply(DeleteEntries param)
 849                         throws Exception {
 850                     //FIXME : Doing nothing for now
 851                     dataSize = 0;
 852                     for (ReplicatedLogEntry entry : journal) {
 853                         dataSize += entry.size();
 854                     }
 855                 }
 856             });
 857         }
 858
 859         @Override public void appendAndPersist(
 860             final ReplicatedLogEntry replicatedLogEntry) {
 861             appendAndPersist(replicatedLogEntry, null);
 862         }
 863
 864         public void appendAndPersist(
 865             final ReplicatedLogEntry replicatedLogEntry,
 866             final Procedure<ReplicatedLogEntry> callback)  {
 867
 868             if(LOG.isDebugEnabled()) {
 869                 LOG.debug("{}: Append log entry and persist {} ", persistenceId(), replicatedLogEntry);
 870             }
 871
 872             // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
 873             journal.add(replicatedLogEntry);
 874
 875             // When persisting events with persist it is guaranteed that the
 876             // persistent actor will not receive further commands between the
 877             // persist call and the execution(s) of the associated event
 878             // handler. This also holds for multiple persist calls in context
 879             // of a single command.
 880             persistence().persist(replicatedLogEntry,
 881                 new Procedure<ReplicatedLogEntry>() {
 882                     @Override
 883                     public void apply(ReplicatedLogEntry evt) throws Exception {
 884                         int logEntrySize = replicatedLogEntry.size();
 885
 886                         dataSize += logEntrySize;
 887                         long dataSizeForCheck = dataSize;
 888
 889                         dataSizeSinceLastSnapshot += logEntrySize;
 890                         long journalSize = lastIndex() + 1;
 891
 892                         if(!hasFollowers()) {
 893                             // When we do not have followers we do not maintain an in-memory log
 894                             // due to this the journalSize will never become anything close to the
 895                             // snapshot batch count. In fact will mostly be 1.
 896                             // Similarly since the journal's dataSize depends on the entries in the
 897                             // journal the journal's dataSize will never reach a value close to the
 898                             // memory threshold.
 899                             // By maintaining the dataSize outside the journal we are tracking essentially
 900                             // what we have written to the disk however since we no longer are in
 901                             // need of doing a snapshot just for the sake of freeing up memory we adjust
 902                             // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
 903                             // as if we were maintaining a real snapshot
 904                             dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
 905                         }
 906
 907                         long dataThreshold = getTotalMemory() *
 908                                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 909
 910                         // when a snaphsot is being taken, captureSnapshot != null
 911                         if (!context.isSnapshotCaptureInitiated() &&
 912                                 ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
 913                                         dataSizeForCheck > dataThreshold)) {
 914
 915                             dataSizeSinceLastSnapshot = 0;
 916
 917                             LOG.info("{}: Initiating Snapshot Capture, journalSize = {}, dataSizeForCheck = {}," +
 918                                 " dataThreshold = {}", persistenceId(), journalSize, dataSizeForCheck, dataThreshold);
 919
 920                             long lastAppliedIndex = -1;
 921                             long lastAppliedTerm = -1;
 922
 923                             ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
 924                             if (!hasFollowers()) {
 925                                 lastAppliedIndex = replicatedLogEntry.getIndex();
 926                                 lastAppliedTerm = replicatedLogEntry.getTerm();
 927                             } else if (lastAppliedEntry != null) {
 928                                 lastAppliedIndex = lastAppliedEntry.getIndex();
 929                                 lastAppliedTerm = lastAppliedEntry.getTerm();
 930                             }
 931
 932                             if(LOG.isDebugEnabled()) {
 933                                 LOG.debug("{}: Snapshot Capture logSize: {}", persistenceId(), journal.size());
 934                                 LOG.debug("{}: Snapshot Capture lastApplied:{} ",
 935                                         persistenceId(), context.getLastApplied());
 936                                 LOG.debug("{}: Snapshot Capture lastAppliedIndex:{}", persistenceId(),
 937                                         lastAppliedIndex);
 938                                 LOG.debug("{}: Snapshot Capture lastAppliedTerm:{}", persistenceId(),
 939                                         lastAppliedTerm);
 940                             }
 941
 942                             // send a CaptureSnapshot to self to make the expensive operation async.
 943                             long replicatedToAllIndex = getCurrentBehavior().getReplicatedToAllIndex();
 944                             ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
 945                             getSelf().tell(new CaptureSnapshot(lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
 946                                 (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
 947                                 (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1)),
 948                                 null);
 949                             context.setSnapshotCaptureInitiated(true);
 950                         }
 951                         if (callback != null){
 952                             callback.apply(replicatedLogEntry);
 953                         }
 954                     }
 955                 }
 956             );
 957         }
 958
 959     }
 960
 961     static class DeleteEntries implements Serializable {
 962         private static final long serialVersionUID = 1L;
 963         private final int fromIndex;
 964
 965         public DeleteEntries(int fromIndex) {
 966             this.fromIndex = fromIndex;
 967         }
 968
 969         public int getFromIndex() {
 970             return fromIndex;
 971         }
 972     }
 973
 974
 975     private class ElectionTermImpl implements ElectionTerm {
 976         /**
 977          * Identifier of the actor whose election term information this is
 978          */
 979         private long currentTerm = 0;
 980         private String votedFor = null;
 981
 982         @Override
 983         public long getCurrentTerm() {
 984             return currentTerm;
 985         }
 986
 987         @Override
 988         public String getVotedFor() {
 989             return votedFor;
 990         }
 991
 992         @Override public void update(long currentTerm, String votedFor) {
 993             if(LOG.isDebugEnabled()) {
 994                 LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
 995             }
 996             this.currentTerm = currentTerm;
 997             this.votedFor = votedFor;
 998         }
 999
1000         @Override
1001         public void updateAndPersist(long currentTerm, String votedFor){
1002             update(currentTerm, votedFor);
1003             // FIXME : Maybe first persist then update the state
1004             persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
1005
1006                 @Override public void apply(UpdateElectionTerm param)
1007                     throws Exception {
1008
1009                 }
1010             });
1011         }
1012     }
1013
1014     static class UpdateElectionTerm implements Serializable {
1015         private static final long serialVersionUID = 1L;
1016         private final long currentTerm;
1017         private final String votedFor;
1018
1019         public UpdateElectionTerm(long currentTerm, String votedFor) {
1020             this.currentTerm = currentTerm;
1021             this.votedFor = votedFor;
1022         }
1023
1024         public long getCurrentTerm() {
1025             return currentTerm;
1026         }
1027
1028         public String getVotedFor() {
1029             return votedFor;
1030         }
1031     }
1032
1033     protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
1034
1035         public NonPersistentRaftDataProvider(){
1036
1037         }
1038
1039         /**
1040          * The way snapshotting works is,
1041          * <ol>
1042          * <li> RaftActor calls createSnapshot on the Shard
1043          * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
1044          * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
1045          * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
1046          * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
1047          * in SaveSnapshotSuccess.
1048          * </ol>
1049          * @param o
1050          */
1051         @Override
1052         public void saveSnapshot(Object o) {
1053             // Make saving Snapshot successful
1054             commitSnapshot(-1L);
1055         }
1056     }
1057
1058     @VisibleForTesting
1059     void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
1060         currentBehavior = behavior;
1061     }
1062
1063     protected RaftActorBehavior getCurrentBehavior() {
1064         return currentBehavior;
1065     }
1066
1067     private static class BehaviorStateHolder {
1068         private RaftActorBehavior behavior;
1069         private String leaderId;
1070
1071         void init(RaftActorBehavior behavior) {
1072             this.behavior = behavior;
1073             this.leaderId = behavior != null ? behavior.getLeaderId() : null;
1074         }
1075
1076         RaftActorBehavior getBehavior() {
1077             return behavior;
1078         }
1079
1080         String getLeaderId() {
1081             return leaderId;
1082         }
1083     }
1084 }