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     public RaftActor(String id, Map<String, String> peerAddresses) {
 117         this(id, peerAddresses, Optional.<ConfigParams>absent());
 118     }
 119
 120     public RaftActor(String id, Map<String, String> peerAddresses,
 121          Optional<ConfigParams> configParams) {
 122
 123         context = new RaftActorContextImpl(this.getSelf(),
 124             this.getContext(), id, new ElectionTermImpl(),
 125             -1, -1, replicatedLog, peerAddresses,
 126             (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
 127             LOG);
 128     }
 129
 130     private void initRecoveryTimer() {
 131         if(recoveryTimer == null) {
 132             recoveryTimer = new Stopwatch();
 133             recoveryTimer.start();
 134         }
 135     }
 136
 137     @Override
 138     public void preStart() throws Exception {
 139         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 140                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 141
 142         super.preStart();
 143     }
 144
 145     @Override
 146     public void handleRecover(Object message) {
 147         if(persistence().isRecoveryApplicable()) {
 148             if (message instanceof SnapshotOffer) {
 149                 onRecoveredSnapshot((SnapshotOffer) message);
 150             } else if (message instanceof ReplicatedLogEntry) {
 151                 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
 152             } else if (message instanceof ApplyLogEntries) {
 153                 onRecoveredApplyLogEntries((ApplyLogEntries) message);
 154             } else if (message instanceof DeleteEntries) {
 155                 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
 156             } else if (message instanceof UpdateElectionTerm) {
 157                 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
 158                         ((UpdateElectionTerm) message).getVotedFor());
 159             } else if (message instanceof RecoveryCompleted) {
 160                 onRecoveryCompletedMessage();
 161             }
 162         } else {
 163             if (message instanceof RecoveryCompleted) {
 164                 // Delete all the messages from the akka journal so that we do not end up with consistency issues
 165                 // Note I am not using the dataPersistenceProvider and directly using the akka api here
 166                 deleteMessages(lastSequenceNr());
 167
 168                 // Delete all the akka snapshots as they will not be needed
 169                 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
 170
 171                 onRecoveryComplete();
 172
 173                 RaftActorBehavior oldBehavior = currentBehavior;
 174                 currentBehavior = new Follower(context);
 175                 handleBehaviorChange(oldBehavior, currentBehavior);
 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(ByteString.copyFrom(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.snapshotIndex, replicatedLog.snapshotTerm);
 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.snapshotIndex,
 272             replicatedLog.snapshotTerm, replicatedLog.size());
 273
 274         RaftActorBehavior oldBehavior = currentBehavior;
 275         currentBehavior = new Follower(context);
 276         handleBehaviorChange(oldBehavior, currentBehavior);
 277     }
 278
 279     @Override public void handleCommand(Object message) {
 280         if (message instanceof ApplyState){
 281             ApplyState applyState = (ApplyState) message;
 282
 283             if(LOG.isDebugEnabled()) {
 284                 LOG.debug("Applying state for log index {} data {}",
 285                     applyState.getReplicatedLogEntry().getIndex(),
 286                     applyState.getReplicatedLogEntry().getData());
 287             }
 288
 289             applyState(applyState.getClientActor(), applyState.getIdentifier(),
 290                 applyState.getReplicatedLogEntry().getData());
 291
 292         } else if (message instanceof ApplyLogEntries){
 293             ApplyLogEntries ale = (ApplyLogEntries) message;
 294             if(LOG.isDebugEnabled()) {
 295                 LOG.debug("Persisting ApplyLogEntries with index={}", ale.getToIndex());
 296             }
 297             persistence().persist(new ApplyLogEntries(ale.getToIndex()), new Procedure<ApplyLogEntries>() {
 298                 @Override
 299                 public void apply(ApplyLogEntries param) throws Exception {
 300                 }
 301             });
 302
 303         } else if(message instanceof ApplySnapshot ) {
 304             Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
 305
 306             if(LOG.isDebugEnabled()) {
 307                 LOG.debug("ApplySnapshot called on Follower Actor " +
 308                         "snapshotIndex:{}, snapshotTerm:{}", snapshot.getLastAppliedIndex(),
 309                     snapshot.getLastAppliedTerm()
 310                 );
 311             }
 312             applySnapshot(ByteString.copyFrom(snapshot.getState()));
 313
 314             //clears the followers log, sets the snapshot index to ensure adjusted-index works
 315             replicatedLog = new ReplicatedLogImpl(snapshot);
 316             context.setReplicatedLog(replicatedLog);
 317             context.setLastApplied(snapshot.getLastAppliedIndex());
 318
 319         } else if (message instanceof FindLeader) {
 320             getSender().tell(
 321                 new FindLeaderReply(getLeaderAddress()),
 322                 getSelf()
 323             );
 324
 325         } else if (message instanceof SaveSnapshotSuccess) {
 326             SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
 327             LOG.info("SaveSnapshotSuccess received for snapshot");
 328
 329             long sequenceNumber = success.metadata().sequenceNr();
 330
 331             commitSnapshot(sequenceNumber);
 332
 333         } else if (message instanceof SaveSnapshotFailure) {
 334             SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
 335
 336             LOG.info("saveSnapshotFailure.metadata():{}", saveSnapshotFailure.metadata().toString());
 337             LOG.error(saveSnapshotFailure.cause(), "SaveSnapshotFailure received for snapshot Cause:");
 338
 339             context.getReplicatedLog().snapshotRollback();
 340
 341             LOG.info("Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
 342                 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}",
 343                 context.getReplicatedLog().getSnapshotIndex(),
 344                 context.getReplicatedLog().getSnapshotTerm(),
 345                 context.getReplicatedLog().size());
 346
 347         } else if (message instanceof CaptureSnapshot) {
 348             LOG.info("CaptureSnapshot received by actor");
 349             CaptureSnapshot cs = (CaptureSnapshot)message;
 350             captureSnapshot = cs;
 351             createSnapshot();
 352
 353         } else if (message instanceof CaptureSnapshotReply){
 354             LOG.info("CaptureSnapshotReply received by actor");
 355             CaptureSnapshotReply csr = (CaptureSnapshotReply) message;
 356
 357             ByteString stateInBytes = csr.getSnapshot();
 358             LOG.info("CaptureSnapshotReply stateInBytes size:{}", stateInBytes.size());
 359             handleCaptureSnapshotReply(stateInBytes);
 360
 361         } else {
 362             if (!(message instanceof AppendEntriesMessages.AppendEntries)
 363                 && !(message instanceof AppendEntriesReply) && !(message instanceof SendHeartBeat)) {
 364                 if(LOG.isDebugEnabled()) {
 365                     LOG.debug("onReceiveCommand: message: {}", message.getClass());
 366                 }
 367             }
 368
 369             RaftActorBehavior oldBehavior = currentBehavior;
 370             currentBehavior = currentBehavior.handleMessage(getSender(), message);
 371
 372             handleBehaviorChange(oldBehavior, currentBehavior);
 373         }
 374     }
 375
 376     private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) {
 377         if (oldBehavior != currentBehavior){
 378             onStateChanged();
 379         }
 380         if (oldBehavior != null) {
 381             // it can happen that the state has not changed but the leader has changed.
 382             onLeaderChanged(oldBehavior.getLeaderId(), currentBehavior.getLeaderId());
 383
 384             if (getRoleChangeNotifier().isPresent() && oldBehavior.state() != currentBehavior.state()) {
 385                 // we do not want to notify when the behavior/role is set for the first time (i.e follower)
 386                 getRoleChangeNotifier().get().tell(new RoleChanged(getId(), oldBehavior.state().name(),
 387                     currentBehavior.state().name()), getSelf());
 388             }
 389         }
 390     }
 391
 392     /**
 393      * When a derived RaftActor needs to persist something it must call
 394      * persistData.
 395      *
 396      * @param clientActor
 397      * @param identifier
 398      * @param data
 399      */
 400     protected void persistData(ActorRef clientActor, String identifier,
 401         Payload data) {
 402
 403         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 404             context.getReplicatedLog().lastIndex() + 1,
 405             context.getTermInformation().getCurrentTerm(), data);
 406
 407         if(LOG.isDebugEnabled()) {
 408             LOG.debug("Persist data {}", replicatedLogEntry);
 409         }
 410
 411         replicatedLog
 412             .appendAndPersist(clientActor, identifier, replicatedLogEntry);
 413     }
 414
 415     protected String getId() {
 416         return context.getId();
 417     }
 418
 419     /**
 420      * Derived actors can call the isLeader method to check if the current
 421      * RaftActor is the Leader or not
 422      *
 423      * @return true it this RaftActor is a Leader false otherwise
 424      */
 425     protected boolean isLeader() {
 426         return context.getId().equals(currentBehavior.getLeaderId());
 427     }
 428
 429     /**
 430      * Derived actor can call getLeader if they need a reference to the Leader.
 431      * This would be useful for example in forwarding a request to an actor
 432      * which is the leader
 433      *
 434      * @return A reference to the leader if known, null otherwise
 435      */
 436     protected ActorSelection getLeader(){
 437         String leaderAddress = getLeaderAddress();
 438
 439         if(leaderAddress == null){
 440             return null;
 441         }
 442
 443         return context.actorSelection(leaderAddress);
 444     }
 445
 446     /**
 447      *
 448      * @return the current leader's id
 449      */
 450     protected String getLeaderId(){
 451         return currentBehavior.getLeaderId();
 452     }
 453
 454     protected RaftState getRaftState() {
 455         return currentBehavior.state();
 456     }
 457
 458     protected ReplicatedLogEntry getLastLogEntry() {
 459         return replicatedLog.last();
 460     }
 461
 462     protected Long getCurrentTerm(){
 463         return context.getTermInformation().getCurrentTerm();
 464     }
 465
 466     protected Long getCommitIndex(){
 467         return context.getCommitIndex();
 468     }
 469
 470     protected Long getLastApplied(){
 471         return context.getLastApplied();
 472     }
 473
 474     protected RaftActorContext getRaftActorContext() {
 475         return context;
 476     }
 477
 478     /**
 479      * setPeerAddress sets the address of a known peer at a later time.
 480      * <p>
 481      * This is to account for situations where a we know that a peer
 482      * exists but we do not know an address up-front. This may also be used in
 483      * situations where a known peer starts off in a different location and we
 484      * need to change it's address
 485      * <p>
 486      * Note that if the peerId does not match the list of peers passed to
 487      * this actor during construction an IllegalStateException will be thrown.
 488      *
 489      * @param peerId
 490      * @param peerAddress
 491      */
 492     protected void setPeerAddress(String peerId, String peerAddress){
 493         context.setPeerAddress(peerId, peerAddress);
 494     }
 495
 496     protected void commitSnapshot(long sequenceNumber) {
 497         context.getReplicatedLog().snapshotCommit();
 498
 499         // TODO: Not sure if we want to be this aggressive with trimming stuff
 500         trimPersistentData(sequenceNumber);
 501     }
 502
 503     /**
 504      * The applyState method will be called by the RaftActor when some data
 505      * needs to be applied to the actor's state
 506      *
 507      * @param clientActor A reference to the client who sent this message. This
 508      *                    is the same reference that was passed to persistData
 509      *                    by the derived actor. clientActor may be null when
 510      *                    the RaftActor is behaving as a follower or during
 511      *                    recovery.
 512      * @param identifier  The identifier of the persisted data. This is also
 513      *                    the same identifier that was passed to persistData by
 514      *                    the derived actor. identifier may be null when
 515      *                    the RaftActor is behaving as a follower or during
 516      *                    recovery
 517      * @param data        A piece of data that was persisted by the persistData call.
 518      *                    This should NEVER be null.
 519      */
 520     protected abstract void applyState(ActorRef clientActor, String identifier,
 521         Object data);
 522
 523     /**
 524      * This method is called during recovery at the start of a batch of state entries. Derived
 525      * classes should perform any initialization needed to start a batch.
 526      */
 527     protected abstract void startLogRecoveryBatch(int maxBatchSize);
 528
 529     /**
 530      * This method is called during recovery to append state data to the current batch. This method
 531      * is called 1 or more times after {@link #startLogRecoveryBatch}.
 532      *
 533      * @param data the state data
 534      */
 535     protected abstract void appendRecoveredLogEntry(Payload data);
 536
 537     /**
 538      * This method is called during recovery to reconstruct the state of the actor.
 539      *
 540      * @param snapshot A snapshot of the state of the actor
 541      */
 542     protected abstract void applyRecoverySnapshot(ByteString snapshot);
 543
 544     /**
 545      * This method is called during recovery at the end of a batch to apply the current batched
 546      * log entries. This method is called after {@link #appendRecoveredLogEntry}.
 547      */
 548     protected abstract void applyCurrentLogRecoveryBatch();
 549
 550     /**
 551      * This method is called when recovery is complete.
 552      */
 553     protected abstract void onRecoveryComplete();
 554
 555     /**
 556      * This method will be called by the RaftActor when a snapshot needs to be
 557      * created. The derived actor should respond with its current state.
 558      * <p/>
 559      * During recovery the state that is returned by the derived actor will
 560      * be passed back to it by calling the applySnapshot  method
 561      *
 562      * @return The current state of the actor
 563      */
 564     protected abstract void createSnapshot();
 565
 566     /**
 567      * This method can be called at any other point during normal
 568      * operations when the derived actor is out of sync with it's peers
 569      * and the only way to bring it in sync is by applying a snapshot
 570      *
 571      * @param snapshot A snapshot of the state of the actor
 572      */
 573     protected abstract void applySnapshot(ByteString snapshot);
 574
 575     /**
 576      * This method will be called by the RaftActor when the state of the
 577      * RaftActor changes. The derived actor can then use methods like
 578      * isLeader or getLeader to do something useful
 579      */
 580     protected abstract void onStateChanged();
 581
 582     protected abstract DataPersistenceProvider persistence();
 583
 584     /**
 585      * Notifier Actor for this RaftActor to notify when a role change happens
 586      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 587      */
 588     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 589
 590     protected void onLeaderChanged(String oldLeader, String newLeader){};
 591
 592     private void trimPersistentData(long sequenceNumber) {
 593         // Trim akka snapshots
 594         // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
 595         // For now guessing that it is ANDed.
 596         persistence().deleteSnapshots(new SnapshotSelectionCriteria(
 597             sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
 598
 599         // Trim akka journal
 600         persistence().deleteMessages(sequenceNumber);
 601     }
 602
 603     private String getLeaderAddress(){
 604         if(isLeader()){
 605             return getSelf().path().toString();
 606         }
 607         String leaderId = currentBehavior.getLeaderId();
 608         if (leaderId == null) {
 609             return null;
 610         }
 611         String peerAddress = context.getPeerAddress(leaderId);
 612         if(LOG.isDebugEnabled()) {
 613             LOG.debug("getLeaderAddress leaderId = {} peerAddress = {}",
 614                     leaderId, peerAddress);
 615         }
 616
 617         return peerAddress;
 618     }
 619
 620     private void handleCaptureSnapshotReply(ByteString stateInBytes) {
 621         // create a snapshot object from the state provided and save it
 622         // when snapshot is saved async, SaveSnapshotSuccess is raised.
 623
 624         Snapshot sn = Snapshot.create(stateInBytes.toByteArray(),
 625             context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
 626             captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
 627             captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
 628
 629         persistence().saveSnapshot(sn);
 630
 631         LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
 632
 633         //be greedy and remove entries from in-mem journal which are in the snapshot
 634         // and update snapshotIndex and snapshotTerm without waiting for the success,
 635
 636         context.getReplicatedLog().snapshotPreCommit(
 637             captureSnapshot.getLastAppliedIndex(),
 638             captureSnapshot.getLastAppliedTerm());
 639
 640         LOG.info("Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
 641             "and term:{}", captureSnapshot.getLastAppliedIndex(),
 642             captureSnapshot.getLastAppliedTerm());
 643
 644         if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
 645             // this would be call straight to the leader and won't initiate in serialization
 646             currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(stateInBytes));
 647         }
 648
 649         captureSnapshot = null;
 650         hasSnapshotCaptureInitiated = false;
 651     }
 652
 653     private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
 654
 655         public ReplicatedLogImpl(Snapshot snapshot) {
 656             super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
 657                 snapshot.getUnAppliedEntries());
 658         }
 659
 660         public ReplicatedLogImpl() {
 661             super();
 662         }
 663
 664         @Override public void removeFromAndPersist(long logEntryIndex) {
 665             int adjustedIndex = adjustedIndex(logEntryIndex);
 666
 667             if (adjustedIndex < 0) {
 668                 return;
 669             }
 670
 671             // FIXME: Maybe this should be done after the command is saved
 672             journal.subList(adjustedIndex , journal.size()).clear();
 673
 674             persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
 675
 676                 @Override public void apply(DeleteEntries param)
 677                     throws Exception {
 678                     //FIXME : Doing nothing for now
 679                     dataSize = 0;
 680                     for(ReplicatedLogEntry entry : journal){
 681                         dataSize += entry.size();
 682                     }
 683                 }
 684             });
 685         }
 686
 687         @Override public void appendAndPersist(
 688             final ReplicatedLogEntry replicatedLogEntry) {
 689             appendAndPersist(null, null, replicatedLogEntry);
 690         }
 691
 692         @Override
 693         public int dataSize() {
 694             return dataSize;
 695         }
 696
 697         public void appendAndPersist(final ActorRef clientActor,
 698             final String identifier,
 699             final ReplicatedLogEntry replicatedLogEntry) {
 700
 701             if(LOG.isDebugEnabled()) {
 702                 LOG.debug("Append log entry and persist {} ", replicatedLogEntry);
 703             }
 704
 705             // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
 706             journal.add(replicatedLogEntry);
 707
 708             // When persisting events with persist it is guaranteed that the
 709             // persistent actor will not receive further commands between the
 710             // persist call and the execution(s) of the associated event
 711             // handler. This also holds for multiple persist calls in context
 712             // of a single command.
 713             persistence().persist(replicatedLogEntry,
 714                 new Procedure<ReplicatedLogEntry>() {
 715                     @Override
 716                     public void apply(ReplicatedLogEntry evt) throws Exception {
 717                         dataSize += replicatedLogEntry.size();
 718
 719                         long dataThreshold = Runtime.getRuntime().totalMemory() *
 720                                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 721
 722                         // when a snaphsot is being taken, captureSnapshot != null
 723                         if (hasSnapshotCaptureInitiated == false &&
 724                                 ( journal.size() % context.getConfigParams().getSnapshotBatchCount() == 0 ||
 725                                         dataSize > dataThreshold)) {
 726
 727                             LOG.info("Initiating Snapshot Capture..");
 728                             long lastAppliedIndex = -1;
 729                             long lastAppliedTerm = -1;
 730
 731                             ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
 732                             if (lastAppliedEntry != null) {
 733                                 lastAppliedIndex = lastAppliedEntry.getIndex();
 734                                 lastAppliedTerm = lastAppliedEntry.getTerm();
 735                             }
 736
 737                             if(LOG.isDebugEnabled()) {
 738                                 LOG.debug("Snapshot Capture logSize: {}", journal.size());
 739                                 LOG.debug("Snapshot Capture lastApplied:{} ",
 740                                     context.getLastApplied());
 741                                 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
 742                                 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
 743                             }
 744
 745                             // send a CaptureSnapshot to self to make the expensive operation async.
 746                             getSelf().tell(new CaptureSnapshot(
 747                                 lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm),
 748                                 null);
 749                             hasSnapshotCaptureInitiated = true;
 750                         }
 751                         // Send message for replication
 752                         if (clientActor != null) {
 753                             currentBehavior.handleMessage(getSelf(),
 754                                 new Replicate(clientActor, identifier,
 755                                     replicatedLogEntry)
 756                             );
 757                         }
 758                     }
 759                 }
 760             );
 761         }
 762
 763     }
 764
 765     static class DeleteEntries implements Serializable {
 766         private static final long serialVersionUID = 1L;
 767         private final int fromIndex;
 768
 769         public DeleteEntries(int fromIndex) {
 770             this.fromIndex = fromIndex;
 771         }
 772
 773         public int getFromIndex() {
 774             return fromIndex;
 775         }
 776     }
 777
 778
 779     private class ElectionTermImpl implements ElectionTerm {
 780         /**
 781          * Identifier of the actor whose election term information this is
 782          */
 783         private long currentTerm = 0;
 784         private String votedFor = null;
 785
 786         @Override
 787         public long getCurrentTerm() {
 788             return currentTerm;
 789         }
 790
 791         @Override
 792         public String getVotedFor() {
 793             return votedFor;
 794         }
 795
 796         @Override public void update(long currentTerm, String votedFor) {
 797             if(LOG.isDebugEnabled()) {
 798                 LOG.debug("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
 799             }
 800             this.currentTerm = currentTerm;
 801             this.votedFor = votedFor;
 802         }
 803
 804         @Override
 805         public void updateAndPersist(long currentTerm, String votedFor){
 806             update(currentTerm, votedFor);
 807             // FIXME : Maybe first persist then update the state
 808             persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
 809
 810                 @Override public void apply(UpdateElectionTerm param)
 811                     throws Exception {
 812
 813                 }
 814             });
 815         }
 816     }
 817
 818     static class UpdateElectionTerm implements Serializable {
 819         private static final long serialVersionUID = 1L;
 820         private final long currentTerm;
 821         private final String votedFor;
 822
 823         public UpdateElectionTerm(long currentTerm, String votedFor) {
 824             this.currentTerm = currentTerm;
 825             this.votedFor = votedFor;
 826         }
 827
 828         public long getCurrentTerm() {
 829             return currentTerm;
 830         }
 831
 832         public String getVotedFor() {
 833             return votedFor;
 834         }
 835     }
 836
 837     protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
 838
 839         public NonPersistentRaftDataProvider(){
 840
 841         }
 842
 843         /**
 844          * The way snapshotting works is,
 845          * <ol>
 846          * <li> RaftActor calls createSnapshot on the Shard
 847          * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 848          * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
 849          * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
 850          * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
 851          * in SaveSnapshotSuccess.
 852          * </ol>
 853          * @param o
 854          */
 855         @Override
 856         public void saveSnapshot(Object o) {
 857             // Make saving Snapshot successful
 858             commitSnapshot(-1L);
 859         }
 860     }
 861
 862     @VisibleForTesting
 863     void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
 864         currentBehavior = behavior;
 865     }
 866
 867     protected RaftActorBehavior getCurrentBehavior() {
 868         return currentBehavior;
 869     }
 870
 871 }