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