opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java

   1 /*
   2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
   3  *
   4  * This program and the accompanying materials are made available under the
   5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   6  * and is available at http://www.eclipse.org/legal/epl-v10.html
   7  */
   8
   9 package org.opendaylight.controller.cluster.raft;
  10
  11 import akka.actor.ActorRef;
  12 import akka.actor.ActorSelection;
  13 import akka.japi.Procedure;
  14 import akka.persistence.RecoveryCompleted;
  15 import akka.persistence.SaveSnapshotFailure;
  16 import akka.persistence.SaveSnapshotSuccess;
  17 import akka.persistence.SnapshotOffer;
  18 import akka.persistence.SnapshotSelectionCriteria;
  19 import com.google.common.annotations.VisibleForTesting;
  20 import com.google.common.base.Optional;
  21 import com.google.common.base.Stopwatch;
  22 import com.google.protobuf.ByteString;
  23 import java.io.Serializable;
  24 import java.util.Map;
  25 import java.util.concurrent.TimeUnit;
  26 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  27 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
  28 import org.opendaylight.controller.cluster.notifications.RoleChanged;
  29 import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
  30 import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
  31 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
  32 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  33 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
  34 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
  35 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  36 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  37 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  38 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  39 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  40 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  41 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  42 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  43 import org.slf4j.Logger;
  44 import org.slf4j.LoggerFactory;
  45
  46 /**
  47  * RaftActor encapsulates a state machine that needs to be kept synchronized
  48  * in a cluster. It implements the RAFT algorithm as described in the paper
  49  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  50  * In Search of an Understandable Consensus Algorithm</a>
  51  * <p/>
  52  * RaftActor has 3 states and each state has a certain behavior associated
  53  * with it. A Raft actor can behave as,
  54  * <ul>
  55  * <li> A Leader </li>
  56  * <li> A Follower (or) </li>
  57  * <li> A Candidate </li>
  58  * </ul>
  59  * <p/>
  60  * <p/>
  61  * A RaftActor MUST be a Leader in order to accept requests from clients to
  62  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  63  * a Leader it is also responsible for ensuring that all followers ultimately
  64  * have the same log and therefore the same state machine as itself.
  65  * <p/>
  66  * <p/>
  67  * The current behavior of a RaftActor determines how election for leadership
  68  * is initiated and how peer RaftActors react to request for votes.
  69  * <p/>
  70  * <p/>
  71  * Each RaftActor also needs to know the current election term. It uses this
  72  * information for a couple of things. One is to simply figure out who it
  73  * voted for in the last election. Another is to figure out if the message
  74  * it received to update it's state is stale.
  75  * <p/>
  76  * <p/>
  77  * The RaftActor uses akka-persistence to store it's replicated log.
  78  * Furthermore through it's behaviors a Raft Actor determines
  79  * <p/>
  80  * <ul>
  81  * <li> when a log entry should be persisted </li>
  82  * <li> when a log entry should be applied to the state machine (and) </li>
  83  * <li> when a snapshot should be saved </li>
  84  * </ul>
  85  */
  86 public abstract class RaftActor extends AbstractUntypedPersistentActor {
  87
  88     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
  89
  90     protected final Logger LOG = LoggerFactory.getLogger(getClass());
  91
  92     /**
  93      * The current state determines the current behavior of a RaftActor
  94      * A Raft Actor always starts off in the Follower State
  95      */
  96     private RaftActorBehavior currentBehavior;
  97
  98     /**
  99      * This context should NOT be passed directly to any other actor it is
 100      * only to be consumed by the RaftActorBehaviors
 101      */
 102     private final RaftActorContext context;
 103
 104     /**
 105      * The in-memory journal
 106      */
 107     private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
 108
 109     private CaptureSnapshot captureSnapshot = null;
 110
 111     private Stopwatch recoveryTimer;
 112
 113     private int currentRecoveryBatchCount;
 114
 115     public RaftActor(String id, Map<String, String> peerAddresses) {
 116         this(id, peerAddresses, Optional.<ConfigParams>absent());
 117     }
 118
 119     public RaftActor(String id, Map<String, String> peerAddresses,
 120          Optional<ConfigParams> configParams) {
 121
 122         context = new RaftActorContextImpl(this.getSelf(),
 123             this.getContext(), id, new ElectionTermImpl(),
 124             -1, -1, replicatedLog, peerAddresses,
 125             (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
 126             LOG);
 127     }
 128
 129     private void initRecoveryTimer() {
 130         if(recoveryTimer == null) {
 131             recoveryTimer = Stopwatch.createStarted();
 132         }
 133     }
 134
 135     @Override
 136     public void preStart() throws Exception {
 137         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 138                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 139
 140         super.preStart();
 141     }
 142
 143     @Override
 144     public void handleRecover(Object message) {
 145         if(persistence().isRecoveryApplicable()) {
 146             if (message instanceof SnapshotOffer) {
 147                 onRecoveredSnapshot((SnapshotOffer) message);
 148             } else if (message instanceof ReplicatedLogEntry) {
 149                 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
 150             } else if (message instanceof ApplyLogEntries) {
 151                 // Handle this message for backwards compatibility with pre-Lithium versions.
 152                 onRecoveredApplyLogEntries(((ApplyLogEntries) message).getToIndex());
 153             } else if (message instanceof ApplyJournalEntries) {
 154                 onRecoveredApplyLogEntries(((ApplyJournalEntries) message).getToIndex());
 155             } else if (message instanceof DeleteEntries) {
 156                 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
 157             } else if (message instanceof UpdateElectionTerm) {
 158                 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
 159                         ((UpdateElectionTerm) message).getVotedFor());
 160             } else if (message instanceof RecoveryCompleted) {
 161                 onRecoveryCompletedMessage();
 162             }
 163         } else {
 164             if (message instanceof RecoveryCompleted) {
 165                 // Delete all the messages from the akka journal so that we do not end up with consistency issues
 166                 // Note I am not using the dataPersistenceProvider and directly using the akka api here
 167                 deleteMessages(lastSequenceNr());
 168
 169                 // Delete all the akka snapshots as they will not be needed
 170                 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
 171
 172                 onRecoveryComplete();
 173
 174                 initializeBehavior();
 175             }
 176         }
 177     }
 178
 179     private void onRecoveredSnapshot(SnapshotOffer offer) {
 180         if(LOG.isDebugEnabled()) {
 181             LOG.debug("{}: SnapshotOffer called..", persistenceId());
 182         }
 183
 184         initRecoveryTimer();
 185
 186         Snapshot snapshot = (Snapshot) offer.snapshot();
 187
 188         // Create a replicated log with the snapshot information
 189         // The replicated log can be used later on to retrieve this snapshot
 190         // when we need to install it on a peer
 191         replicatedLog = new ReplicatedLogImpl(snapshot);
 192
 193         context.setReplicatedLog(replicatedLog);
 194         context.setLastApplied(snapshot.getLastAppliedIndex());
 195         context.setCommitIndex(snapshot.getLastAppliedIndex());
 196
 197         Stopwatch timer = Stopwatch.createStarted();
 198
 199         // Apply the snapshot to the actors state
 200         applyRecoverySnapshot(snapshot.getState());
 201
 202         timer.stop();
 203         LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
 204                 replicatedLog.size(), persistenceId(), timer.toString(),
 205                 replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm());
 206     }
 207
 208     private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
 209         if(LOG.isDebugEnabled()) {
 210             LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex());
 211         }
 212
 213         replicatedLog.append(logEntry);
 214     }
 215
 216     private void onRecoveredApplyLogEntries(long toIndex) {
 217         if(LOG.isDebugEnabled()) {
 218             LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}",
 219                     persistenceId(), context.getLastApplied() + 1, toIndex);
 220         }
 221
 222         for (long i = context.getLastApplied() + 1; i <= toIndex; i++) {
 223             batchRecoveredLogEntry(replicatedLog.get(i));
 224         }
 225
 226         context.setLastApplied(toIndex);
 227         context.setCommitIndex(toIndex);
 228     }
 229
 230     private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
 231         initRecoveryTimer();
 232
 233         int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
 234         if(currentRecoveryBatchCount == 0) {
 235             startLogRecoveryBatch(batchSize);
 236         }
 237
 238         appendRecoveredLogEntry(logEntry.getData());
 239
 240         if(++currentRecoveryBatchCount >= batchSize) {
 241             endCurrentLogRecoveryBatch();
 242         }
 243     }
 244
 245     private void endCurrentLogRecoveryBatch() {
 246         applyCurrentLogRecoveryBatch();
 247         currentRecoveryBatchCount = 0;
 248     }
 249
 250     private void onRecoveryCompletedMessage() {
 251         if(currentRecoveryBatchCount > 0) {
 252             endCurrentLogRecoveryBatch();
 253         }
 254
 255         onRecoveryComplete();
 256
 257         String recoveryTime = "";
 258         if(recoveryTimer != null) {
 259             recoveryTimer.stop();
 260             recoveryTime = " in " + recoveryTimer.toString();
 261             recoveryTimer = null;
 262         }
 263
 264         LOG.info(
 265             "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
 266                 "Persistence Id =  " + persistenceId() +
 267                 " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
 268                 "journal-size={}",
 269             replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(),
 270             replicatedLog.getSnapshotTerm(), replicatedLog.size());
 271
 272         initializeBehavior();
 273     }
 274
 275     protected void initializeBehavior(){
 276         changeCurrentBehavior(new Follower(context));
 277     }
 278
 279     protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
 280         RaftActorBehavior oldBehavior = currentBehavior;
 281         currentBehavior = newBehavior;
 282         handleBehaviorChange(oldBehavior, currentBehavior);
 283     }
 284
 285     @Override public void handleCommand(Object message) {
 286         if (message instanceof ApplyState){
 287             ApplyState applyState = (ApplyState) message;
 288
 289             long elapsedTime = (System.nanoTime() - applyState.getStartTime());
 290             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 291                 LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 292                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 293             }
 294
 295             if(LOG.isDebugEnabled()) {
 296                 LOG.debug("{}: Applying state for log index {} data {}",
 297                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 298                     applyState.getReplicatedLogEntry().getData());
 299             }
 300
 301             applyState(applyState.getClientActor(), applyState.getIdentifier(),
 302                 applyState.getReplicatedLogEntry().getData());
 303
 304         } else if (message instanceof ApplyJournalEntries){
 305             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 306             if(LOG.isDebugEnabled()) {
 307                 LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), applyEntries.getToIndex());
 308             }
 309             persistence().persist(applyEntries, new Procedure<ApplyJournalEntries>() {
 310                 @Override
 311                 public void apply(ApplyJournalEntries param) throws Exception {
 312                 }
 313             });
 314
 315         } else if(message instanceof ApplySnapshot ) {
 316             Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
 317
 318             if(LOG.isDebugEnabled()) {
 319                 LOG.debug("{}: ApplySnapshot called on Follower Actor " +
 320                         "snapshotIndex:{}, snapshotTerm:{}", persistenceId(), snapshot.getLastAppliedIndex(),
 321                     snapshot.getLastAppliedTerm()
 322                 );
 323             }
 324
 325             applySnapshot(snapshot.getState());
 326
 327             //clears the followers log, sets the snapshot index to ensure adjusted-index works
 328             replicatedLog = new ReplicatedLogImpl(snapshot);
 329             context.setReplicatedLog(replicatedLog);
 330             context.setLastApplied(snapshot.getLastAppliedIndex());
 331
 332         } else if (message instanceof FindLeader) {
 333             getSender().tell(
 334                 new FindLeaderReply(getLeaderAddress()),
 335                 getSelf()
 336             );
 337
 338         } else if (message instanceof SaveSnapshotSuccess) {
 339             SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
 340             LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId());
 341
 342             long sequenceNumber = success.metadata().sequenceNr();
 343
 344             commitSnapshot(sequenceNumber);
 345
 346         } else if (message instanceof SaveSnapshotFailure) {
 347             SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
 348
 349             LOG.error("{}: SaveSnapshotFailure received for snapshot Cause:",
 350                     persistenceId(), saveSnapshotFailure.cause());
 351
 352             context.getReplicatedLog().snapshotRollback();
 353
 354             LOG.info("{}: Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
 355                 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}", persistenceId(),
 356                 context.getReplicatedLog().getSnapshotIndex(),
 357                 context.getReplicatedLog().getSnapshotTerm(),
 358                 context.getReplicatedLog().size());
 359
 360         } else if (message instanceof CaptureSnapshot) {
 361             LOG.info("{}: CaptureSnapshot received by actor", persistenceId());
 362
 363             if(captureSnapshot == null) {
 364                 captureSnapshot = (CaptureSnapshot)message;
 365                 createSnapshot();
 366             }
 367
 368         } else if (message instanceof CaptureSnapshotReply){
 369             handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot());
 370
 371         } else {
 372             RaftActorBehavior oldBehavior = currentBehavior;
 373             currentBehavior = currentBehavior.handleMessage(getSender(), message);
 374
 375             handleBehaviorChange(oldBehavior, currentBehavior);
 376         }
 377     }
 378
 379     private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) {
 380         if (oldBehavior != currentBehavior){
 381             onStateChanged();
 382         }
 383
 384         String oldBehaviorLeaderId = oldBehavior == null? null : oldBehavior.getLeaderId();
 385         String oldBehaviorState = oldBehavior == null? null : oldBehavior.state().name();
 386
 387         // it can happen that the state has not changed but the leader has changed.
 388         onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
 389
 390         if (getRoleChangeNotifier().isPresent() &&
 391                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
 392             getRoleChangeNotifier().get().tell(
 393                     new RoleChanged(getId(), oldBehaviorState , currentBehavior.state().name()),
 394                     getSelf());
 395         }
 396     }
 397
 398     /**
 399      * When a derived RaftActor needs to persist something it must call
 400      * persistData.
 401      *
 402      * @param clientActor
 403      * @param identifier
 404      * @param data
 405      */
 406     protected void persistData(final ActorRef clientActor, final String identifier,
 407         final Payload data) {
 408
 409         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 410             context.getReplicatedLog().lastIndex() + 1,
 411             context.getTermInformation().getCurrentTerm(), data);
 412
 413         if(LOG.isDebugEnabled()) {
 414             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 415         }
 416
 417         final RaftActorContext raftContext = getRaftActorContext();
 418
 419         replicatedLog
 420                 .appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
 421                     @Override
 422                     public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
 423                         if(!hasFollowers()){
 424                             // Increment the Commit Index and the Last Applied values
 425                             raftContext.setCommitIndex(replicatedLogEntry.getIndex());
 426                             raftContext.setLastApplied(replicatedLogEntry.getIndex());
 427
 428                             // Apply the state immediately
 429                             applyState(clientActor, identifier, data);
 430
 431                             // Send a ApplyJournalEntries message so that we write the fact that we applied
 432                             // the state to durable storage
 433                             self().tell(new ApplyJournalEntries(replicatedLogEntry.getIndex()), self());
 434
 435                             // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot
 436                             if(!context.isSnapshotCaptureInitiated()){
 437                                 raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(),
 438                                         raftContext.getTermInformation().getCurrentTerm());
 439                                 raftContext.getReplicatedLog().snapshotCommit();
 440                             } else {
 441                                 LOG.debug("{}: Skipping fake snapshotting for {} because real snapshotting is in progress",
 442                                         persistenceId(), getId());
 443                             }
 444                         } else if (clientActor != null) {
 445                             // Send message for replication
 446                             currentBehavior.handleMessage(getSelf(),
 447                                     new Replicate(clientActor, identifier,
 448                                             replicatedLogEntry)
 449                             );
 450                         }
 451
 452                     }
 453                 });    }
 454
 455     protected String getId() {
 456         return context.getId();
 457     }
 458
 459     /**
 460      * Derived actors can call the isLeader method to check if the current
 461      * RaftActor is the Leader or not
 462      *
 463      * @return true it this RaftActor is a Leader false otherwise
 464      */
 465     protected boolean isLeader() {
 466         return context.getId().equals(currentBehavior.getLeaderId());
 467     }
 468
 469     /**
 470      * Derived actor can call getLeader if they need a reference to the Leader.
 471      * This would be useful for example in forwarding a request to an actor
 472      * which is the leader
 473      *
 474      * @return A reference to the leader if known, null otherwise
 475      */
 476     protected ActorSelection getLeader(){
 477         String leaderAddress = getLeaderAddress();
 478
 479         if(leaderAddress == null){
 480             return null;
 481         }
 482
 483         return context.actorSelection(leaderAddress);
 484     }
 485
 486     /**
 487      *
 488      * @return the current leader's id
 489      */
 490     protected String getLeaderId(){
 491         return currentBehavior.getLeaderId();
 492     }
 493
 494     protected RaftState getRaftState() {
 495         return currentBehavior.state();
 496     }
 497
 498     protected ReplicatedLogEntry getLastLogEntry() {
 499         return replicatedLog.last();
 500     }
 501
 502     protected Long getCurrentTerm(){
 503         return context.getTermInformation().getCurrentTerm();
 504     }
 505
 506     protected Long getCommitIndex(){
 507         return context.getCommitIndex();
 508     }
 509
 510     protected Long getLastApplied(){
 511         return context.getLastApplied();
 512     }
 513
 514     protected RaftActorContext getRaftActorContext() {
 515         return context;
 516     }
 517
 518     /**
 519      * setPeerAddress sets the address of a known peer at a later time.
 520      * <p>
 521      * This is to account for situations where a we know that a peer
 522      * exists but we do not know an address up-front. This may also be used in
 523      * situations where a known peer starts off in a different location and we
 524      * need to change it's address
 525      * <p>
 526      * Note that if the peerId does not match the list of peers passed to
 527      * this actor during construction an IllegalStateException will be thrown.
 528      *
 529      * @param peerId
 530      * @param peerAddress
 531      */
 532     protected void setPeerAddress(String peerId, String peerAddress){
 533         context.setPeerAddress(peerId, peerAddress);
 534     }
 535
 536     protected void commitSnapshot(long sequenceNumber) {
 537         context.getReplicatedLog().snapshotCommit();
 538
 539         // TODO: Not sure if we want to be this aggressive with trimming stuff
 540         trimPersistentData(sequenceNumber);
 541     }
 542
 543     /**
 544      * The applyState method will be called by the RaftActor when some data
 545      * needs to be applied to the actor's state
 546      *
 547      * @param clientActor A reference to the client who sent this message. This
 548      *                    is the same reference that was passed to persistData
 549      *                    by the derived actor. clientActor may be null when
 550      *                    the RaftActor is behaving as a follower or during
 551      *                    recovery.
 552      * @param identifier  The identifier of the persisted data. This is also
 553      *                    the same identifier that was passed to persistData by
 554      *                    the derived actor. identifier may be null when
 555      *                    the RaftActor is behaving as a follower or during
 556      *                    recovery
 557      * @param data        A piece of data that was persisted by the persistData call.
 558      *                    This should NEVER be null.
 559      */
 560     protected abstract void applyState(ActorRef clientActor, String identifier,
 561         Object data);
 562
 563     /**
 564      * This method is called during recovery at the start of a batch of state entries. Derived
 565      * classes should perform any initialization needed to start a batch.
 566      */
 567     protected abstract void startLogRecoveryBatch(int maxBatchSize);
 568
 569     /**
 570      * This method is called during recovery to append state data to the current batch. This method
 571      * is called 1 or more times after {@link #startLogRecoveryBatch}.
 572      *
 573      * @param data the state data
 574      */
 575     protected abstract void appendRecoveredLogEntry(Payload data);
 576
 577     /**
 578      * This method is called during recovery to reconstruct the state of the actor.
 579      *
 580      * @param snapshotBytes A snapshot of the state of the actor
 581      */
 582     protected abstract void applyRecoverySnapshot(byte[] snapshotBytes);
 583
 584     /**
 585      * This method is called during recovery at the end of a batch to apply the current batched
 586      * log entries. This method is called after {@link #appendRecoveredLogEntry}.
 587      */
 588     protected abstract void applyCurrentLogRecoveryBatch();
 589
 590     /**
 591      * This method is called when recovery is complete.
 592      */
 593     protected abstract void onRecoveryComplete();
 594
 595     /**
 596      * This method will be called by the RaftActor when a snapshot needs to be
 597      * created. The derived actor should respond with its current state.
 598      * <p/>
 599      * During recovery the state that is returned by the derived actor will
 600      * be passed back to it by calling the applySnapshot  method
 601      *
 602      * @return The current state of the actor
 603      */
 604     protected abstract void createSnapshot();
 605
 606     /**
 607      * This method can be called at any other point during normal
 608      * operations when the derived actor is out of sync with it's peers
 609      * and the only way to bring it in sync is by applying a snapshot
 610      *
 611      * @param snapshotBytes A snapshot of the state of the actor
 612      */
 613     protected abstract void applySnapshot(byte[] snapshotBytes);
 614
 615     /**
 616      * This method will be called by the RaftActor when the state of the
 617      * RaftActor changes. The derived actor can then use methods like
 618      * isLeader or getLeader to do something useful
 619      */
 620     protected abstract void onStateChanged();
 621
 622     protected abstract DataPersistenceProvider persistence();
 623
 624     /**
 625      * Notifier Actor for this RaftActor to notify when a role change happens
 626      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 627      */
 628     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 629
 630     protected void onLeaderChanged(String oldLeader, String newLeader){};
 631
 632     private void trimPersistentData(long sequenceNumber) {
 633         // Trim akka snapshots
 634         // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
 635         // For now guessing that it is ANDed.
 636         persistence().deleteSnapshots(new SnapshotSelectionCriteria(
 637             sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
 638
 639         // Trim akka journal
 640         persistence().deleteMessages(sequenceNumber);
 641     }
 642
 643     private String getLeaderAddress(){
 644         if(isLeader()){
 645             return getSelf().path().toString();
 646         }
 647         String leaderId = currentBehavior.getLeaderId();
 648         if (leaderId == null) {
 649             return null;
 650         }
 651         String peerAddress = context.getPeerAddress(leaderId);
 652         if(LOG.isDebugEnabled()) {
 653             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 654                     persistenceId(), leaderId, peerAddress);
 655         }
 656
 657         return peerAddress;
 658     }
 659
 660     private void handleCaptureSnapshotReply(byte[] snapshotBytes) {
 661         LOG.info("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length);
 662
 663         // create a snapshot object from the state provided and save it
 664         // when snapshot is saved async, SaveSnapshotSuccess is raised.
 665
 666         Snapshot sn = Snapshot.create(snapshotBytes,
 667             context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
 668             captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
 669             captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
 670
 671         persistence().saveSnapshot(sn);
 672
 673         LOG.info("{}: Persisting of snapshot done:{}", persistenceId(), sn.getLogMessage());
 674
 675         long dataThreshold = Runtime.getRuntime().totalMemory() *
 676                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 677         if (context.getReplicatedLog().dataSize() > dataThreshold) {
 678             // if memory is less, clear the log based on lastApplied.
 679             // this could/should only happen if one of the followers is down
 680             // as normally we keep removing from the log when its replicated to all.
 681             context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getLastAppliedIndex(),
 682                     captureSnapshot.getLastAppliedTerm());
 683
 684             getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
 685         } else if(captureSnapshot.getReplicatedToAllIndex() != -1){
 686             // clear the log based on replicatedToAllIndex
 687             context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getReplicatedToAllIndex(),
 688                     captureSnapshot.getReplicatedToAllTerm());
 689
 690             getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
 691         } else {
 692             // The replicatedToAllIndex was not found in the log
 693             // This means that replicatedToAllIndex never moved beyond -1 or that it is already in the snapshot.
 694             // In this scenario we may need to save the snapshot to the akka persistence
 695             // snapshot for recovery but we do not need to do the replicated log trimming.
 696             context.getReplicatedLog().snapshotPreCommit(replicatedLog.getSnapshotIndex(),
 697                     replicatedLog.getSnapshotTerm());
 698         }
 699
 700
 701         LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
 702             "and term:{}", persistenceId(), captureSnapshot.getLastAppliedIndex(),
 703             captureSnapshot.getLastAppliedTerm());
 704
 705         if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
 706             // this would be call straight to the leader and won't initiate in serialization
 707             currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(
 708                     ByteString.copyFrom(snapshotBytes)));
 709         }
 710
 711         captureSnapshot = null;
 712         context.setSnapshotCaptureInitiated(false);
 713     }
 714
 715     protected boolean hasFollowers(){
 716         return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
 717     }
 718
 719     private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
 720
 721         private static final int DATA_SIZE_DIVIDER = 5;
 722         private long dataSizeSinceLastSnapshot = 0;
 723
 724         public ReplicatedLogImpl(Snapshot snapshot) {
 725             super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
 726                 snapshot.getUnAppliedEntries());
 727         }
 728
 729         public ReplicatedLogImpl() {
 730             super();
 731         }
 732
 733         @Override public void removeFromAndPersist(long logEntryIndex) {
 734             int adjustedIndex = adjustedIndex(logEntryIndex);
 735
 736             if (adjustedIndex < 0) {
 737                 return;
 738             }
 739
 740             // FIXME: Maybe this should be done after the command is saved
 741             journal.subList(adjustedIndex , journal.size()).clear();
 742
 743             persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>() {
 744
 745                 @Override
 746                 public void apply(DeleteEntries param)
 747                         throws Exception {
 748                     //FIXME : Doing nothing for now
 749                     dataSize = 0;
 750                     for (ReplicatedLogEntry entry : journal) {
 751                         dataSize += entry.size();
 752                     }
 753                 }
 754             });
 755         }
 756
 757         @Override public void appendAndPersist(
 758             final ReplicatedLogEntry replicatedLogEntry) {
 759             appendAndPersist(replicatedLogEntry, null);
 760         }
 761
 762         public void appendAndPersist(
 763             final ReplicatedLogEntry replicatedLogEntry,
 764             final Procedure<ReplicatedLogEntry> callback)  {
 765
 766             if(LOG.isDebugEnabled()) {
 767                 LOG.debug("{}: Append log entry and persist {} ", persistenceId(), replicatedLogEntry);
 768             }
 769
 770             // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
 771             journal.add(replicatedLogEntry);
 772
 773             // When persisting events with persist it is guaranteed that the
 774             // persistent actor will not receive further commands between the
 775             // persist call and the execution(s) of the associated event
 776             // handler. This also holds for multiple persist calls in context
 777             // of a single command.
 778             persistence().persist(replicatedLogEntry,
 779                 new Procedure<ReplicatedLogEntry>() {
 780                     @Override
 781                     public void apply(ReplicatedLogEntry evt) throws Exception {
 782                         int logEntrySize = replicatedLogEntry.size();
 783
 784                         dataSize += logEntrySize;
 785                         long dataSizeForCheck = dataSize;
 786
 787                         dataSizeSinceLastSnapshot += logEntrySize;
 788                         long journalSize = lastIndex() + 1;
 789
 790                         if(!hasFollowers()) {
 791                             // When we do not have followers we do not maintain an in-memory log
 792                             // due to this the journalSize will never become anything close to the
 793                             // snapshot batch count. In fact will mostly be 1.
 794                             // Similarly since the journal's dataSize depends on the entries in the
 795                             // journal the journal's dataSize will never reach a value close to the
 796                             // memory threshold.
 797                             // By maintaining the dataSize outside the journal we are tracking essentially
 798                             // what we have written to the disk however since we no longer are in
 799                             // need of doing a snapshot just for the sake of freeing up memory we adjust
 800                             // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
 801                             // as if we were maintaining a real snapshot
 802                             dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
 803                         }
 804
 805                         long dataThreshold = Runtime.getRuntime().totalMemory() *
 806                                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 807
 808                         // when a snaphsot is being taken, captureSnapshot != null
 809                         if (!context.isSnapshotCaptureInitiated() &&
 810                                 ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
 811                                         dataSizeForCheck > dataThreshold)) {
 812
 813                             dataSizeSinceLastSnapshot = 0;
 814
 815                             LOG.info("{}: Initiating Snapshot Capture, journalSize = {}, dataSizeForCheck = {}," +
 816                                 " dataThreshold = {}", persistenceId(), journalSize, dataSizeForCheck, dataThreshold);
 817
 818                             long lastAppliedIndex = -1;
 819                             long lastAppliedTerm = -1;
 820
 821                             ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
 822                             if (!hasFollowers()) {
 823                                 lastAppliedIndex = replicatedLogEntry.getIndex();
 824                                 lastAppliedTerm = replicatedLogEntry.getTerm();
 825                             } else if (lastAppliedEntry != null) {
 826                                 lastAppliedIndex = lastAppliedEntry.getIndex();
 827                                 lastAppliedTerm = lastAppliedEntry.getTerm();
 828                             }
 829
 830                             if(LOG.isDebugEnabled()) {
 831                                 LOG.debug("{}: Snapshot Capture logSize: {}", persistenceId(), journal.size());
 832                                 LOG.debug("{}: Snapshot Capture lastApplied:{} ",
 833                                         persistenceId(), context.getLastApplied());
 834                                 LOG.debug("{}: Snapshot Capture lastAppliedIndex:{}", persistenceId(),
 835                                         lastAppliedIndex);
 836                                 LOG.debug("{}: Snapshot Capture lastAppliedTerm:{}", persistenceId(),
 837                                         lastAppliedTerm);
 838                             }
 839
 840                             // send a CaptureSnapshot to self to make the expensive operation async.
 841                             long replicatedToAllIndex = getCurrentBehavior().getReplicatedToAllIndex();
 842                             ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
 843                             getSelf().tell(new CaptureSnapshot(lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
 844                                 (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
 845                                 (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1)),
 846                                 null);
 847                             context.setSnapshotCaptureInitiated(true);
 848                         }
 849                         if (callback != null){
 850                             callback.apply(replicatedLogEntry);
 851                         }
 852                     }
 853                 }
 854             );
 855         }
 856
 857     }
 858
 859     static class DeleteEntries implements Serializable {
 860         private static final long serialVersionUID = 1L;
 861         private final int fromIndex;
 862
 863         public DeleteEntries(int fromIndex) {
 864             this.fromIndex = fromIndex;
 865         }
 866
 867         public int getFromIndex() {
 868             return fromIndex;
 869         }
 870     }
 871
 872
 873     private class ElectionTermImpl implements ElectionTerm {
 874         /**
 875          * Identifier of the actor whose election term information this is
 876          */
 877         private long currentTerm = 0;
 878         private String votedFor = null;
 879
 880         @Override
 881         public long getCurrentTerm() {
 882             return currentTerm;
 883         }
 884
 885         @Override
 886         public String getVotedFor() {
 887             return votedFor;
 888         }
 889
 890         @Override public void update(long currentTerm, String votedFor) {
 891             if(LOG.isDebugEnabled()) {
 892                 LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
 893             }
 894             this.currentTerm = currentTerm;
 895             this.votedFor = votedFor;
 896         }
 897
 898         @Override
 899         public void updateAndPersist(long currentTerm, String votedFor){
 900             update(currentTerm, votedFor);
 901             // FIXME : Maybe first persist then update the state
 902             persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
 903
 904                 @Override public void apply(UpdateElectionTerm param)
 905                     throws Exception {
 906
 907                 }
 908             });
 909         }
 910     }
 911
 912     static class UpdateElectionTerm implements Serializable {
 913         private static final long serialVersionUID = 1L;
 914         private final long currentTerm;
 915         private final String votedFor;
 916
 917         public UpdateElectionTerm(long currentTerm, String votedFor) {
 918             this.currentTerm = currentTerm;
 919             this.votedFor = votedFor;
 920         }
 921
 922         public long getCurrentTerm() {
 923             return currentTerm;
 924         }
 925
 926         public String getVotedFor() {
 927             return votedFor;
 928         }
 929     }
 930
 931     protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
 932
 933         public NonPersistentRaftDataProvider(){
 934
 935         }
 936
 937         /**
 938          * The way snapshotting works is,
 939          * <ol>
 940          * <li> RaftActor calls createSnapshot on the Shard
 941          * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
 942          * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
 943          * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
 944          * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
 945          * in SaveSnapshotSuccess.
 946          * </ol>
 947          * @param o
 948          */
 949         @Override
 950         public void saveSnapshot(Object o) {
 951             // Make saving Snapshot successful
 952             commitSnapshot(-1L);
 953         }
 954     }
 955
 956     @VisibleForTesting
 957     void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
 958         currentBehavior = behavior;
 959     }
 960
 961     protected RaftActorBehavior getCurrentBehavior() {
 962         return currentBehavior;
 963     }
 964
 965 }