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