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