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