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.common.collect.ImmutableMap;
  23 import com.google.common.collect.Lists;
  24 import com.google.protobuf.ByteString;
  25 import java.io.Serializable;
  26 import java.util.Collection;
  27 import java.util.List;
  28 import java.util.Map;
  29 import java.util.concurrent.TimeUnit;
  30 import org.apache.commons.lang3.time.DurationFormatUtils;
  31 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  32 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
  33 import org.opendaylight.controller.cluster.notifications.RoleChanged;
  34 import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
  35 import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
  36 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
  37 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  38 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
  39 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
  40 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  41 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
  43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
  44 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
  45 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
  46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
  47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
  48 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
  49 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
  50 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
  51 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  52 import org.slf4j.Logger;
  53 import org.slf4j.LoggerFactory;
  54
  55 /**
  56  * RaftActor encapsulates a state machine that needs to be kept synchronized
  57  * in a cluster. It implements the RAFT algorithm as described in the paper
  58  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
  59  * In Search of an Understandable Consensus Algorithm</a>
  60  * <p/>
  61  * RaftActor has 3 states and each state has a certain behavior associated
  62  * with it. A Raft actor can behave as,
  63  * <ul>
  64  * <li> A Leader </li>
  65  * <li> A Follower (or) </li>
  66  * <li> A Candidate </li>
  67  * </ul>
  68  * <p/>
  69  * <p/>
  70  * A RaftActor MUST be a Leader in order to accept requests from clients to
  71  * change the state of it's encapsulated state machine. Once a RaftActor becomes
  72  * a Leader it is also responsible for ensuring that all followers ultimately
  73  * have the same log and therefore the same state machine as itself.
  74  * <p/>
  75  * <p/>
  76  * The current behavior of a RaftActor determines how election for leadership
  77  * is initiated and how peer RaftActors react to request for votes.
  78  * <p/>
  79  * <p/>
  80  * Each RaftActor also needs to know the current election term. It uses this
  81  * information for a couple of things. One is to simply figure out who it
  82  * voted for in the last election. Another is to figure out if the message
  83  * it received to update it's state is stale.
  84  * <p/>
  85  * <p/>
  86  * The RaftActor uses akka-persistence to store it's replicated log.
  87  * Furthermore through it's behaviors a Raft Actor determines
  88  * <p/>
  89  * <ul>
  90  * <li> when a log entry should be persisted </li>
  91  * <li> when a log entry should be applied to the state machine (and) </li>
  92  * <li> when a snapshot should be saved </li>
  93  * </ul>
  94  */
  95 public abstract class RaftActor extends AbstractUntypedPersistentActor {
  96
  97     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
  98
  99     private static final Procedure<ApplyJournalEntries> APPLY_JOURNAL_ENTRIES_PERSIST_CALLBACK =
 100             new Procedure<ApplyJournalEntries>() {
 101                 @Override
 102                 public void apply(ApplyJournalEntries param) throws Exception {
 103                 }
 104             };
 105
 106     protected final Logger LOG = LoggerFactory.getLogger(getClass());
 107
 108     /**
 109      * The current state determines the current behavior of a RaftActor
 110      * A Raft Actor always starts off in the Follower State
 111      */
 112     private RaftActorBehavior currentBehavior;
 113
 114     /**
 115      * This context should NOT be passed directly to any other actor it is
 116      * only to be consumed by the RaftActorBehaviors
 117      */
 118     private final RaftActorContextImpl context;
 119
 120     /**
 121      * The in-memory journal
 122      */
 123     private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
 124
 125     private CaptureSnapshot captureSnapshot = null;
 126
 127     private Stopwatch recoveryTimer;
 128
 129     private int currentRecoveryBatchCount;
 130
 131     public RaftActor(String id, Map<String, String> peerAddresses) {
 132         this(id, peerAddresses, Optional.<ConfigParams>absent());
 133     }
 134
 135     public RaftActor(String id, Map<String, String> peerAddresses,
 136          Optional<ConfigParams> configParams) {
 137
 138         context = new RaftActorContextImpl(this.getSelf(),
 139             this.getContext(), id, new ElectionTermImpl(),
 140             -1, -1, replicatedLog, peerAddresses,
 141             (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
 142             LOG);
 143     }
 144
 145     private void initRecoveryTimer() {
 146         if(recoveryTimer == null) {
 147             recoveryTimer = Stopwatch.createStarted();
 148         }
 149     }
 150
 151     @Override
 152     public void preStart() throws Exception {
 153         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
 154                 context.getConfigParams().getJournalRecoveryLogBatchSize());
 155
 156         super.preStart();
 157     }
 158
 159     @Override
 160     public void postStop() {
 161         if(currentBehavior != null) {
 162             try {
 163                 currentBehavior.close();
 164             } catch (Exception e) {
 165                 LOG.debug("{}: Error closing behavior {}", persistenceId(), currentBehavior.state());
 166             }
 167         }
 168
 169         super.postStop();
 170     }
 171
 172     @Override
 173     public void handleRecover(Object message) {
 174         if(persistence().isRecoveryApplicable()) {
 175             if (message instanceof SnapshotOffer) {
 176                 onRecoveredSnapshot((SnapshotOffer) message);
 177             } else if (message instanceof ReplicatedLogEntry) {
 178                 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
 179             } else if (message instanceof ApplyLogEntries) {
 180                 // Handle this message for backwards compatibility with pre-Lithium versions.
 181                 onRecoveredApplyLogEntries(((ApplyLogEntries) message).getToIndex());
 182             } else if (message instanceof ApplyJournalEntries) {
 183                 onRecoveredApplyLogEntries(((ApplyJournalEntries) message).getToIndex());
 184             } else if (message instanceof DeleteEntries) {
 185                 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
 186             } else if (message instanceof UpdateElectionTerm) {
 187                 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
 188                         ((UpdateElectionTerm) message).getVotedFor());
 189             } else if (message instanceof RecoveryCompleted) {
 190                 onRecoveryCompletedMessage();
 191             }
 192         } else {
 193             if (message instanceof RecoveryCompleted) {
 194                 // Delete all the messages from the akka journal so that we do not end up with consistency issues
 195                 // Note I am not using the dataPersistenceProvider and directly using the akka api here
 196                 deleteMessages(lastSequenceNr());
 197
 198                 // Delete all the akka snapshots as they will not be needed
 199                 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
 200
 201                 onRecoveryComplete();
 202
 203                 initializeBehavior();
 204             }
 205         }
 206     }
 207
 208     private void onRecoveredSnapshot(SnapshotOffer offer) {
 209         if(LOG.isDebugEnabled()) {
 210             LOG.debug("{}: SnapshotOffer called..", persistenceId());
 211         }
 212
 213         initRecoveryTimer();
 214
 215         Snapshot snapshot = (Snapshot) offer.snapshot();
 216
 217         // Create a replicated log with the snapshot information
 218         // The replicated log can be used later on to retrieve this snapshot
 219         // when we need to install it on a peer
 220         replicatedLog = new ReplicatedLogImpl(snapshot);
 221
 222         context.setReplicatedLog(replicatedLog);
 223         context.setLastApplied(snapshot.getLastAppliedIndex());
 224         context.setCommitIndex(snapshot.getLastAppliedIndex());
 225
 226         Stopwatch timer = Stopwatch.createStarted();
 227
 228         // Apply the snapshot to the actors state
 229         applyRecoverySnapshot(snapshot.getState());
 230
 231         timer.stop();
 232         LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
 233                 replicatedLog.size(), persistenceId(), timer.toString(),
 234                 replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm());
 235     }
 236
 237     private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
 238         if(LOG.isDebugEnabled()) {
 239             LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex());
 240         }
 241
 242         replicatedLog.append(logEntry);
 243     }
 244
 245     private void onRecoveredApplyLogEntries(long toIndex) {
 246         if(LOG.isDebugEnabled()) {
 247             LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}",
 248                     persistenceId(), context.getLastApplied() + 1, toIndex);
 249         }
 250
 251         for (long i = context.getLastApplied() + 1; i <= toIndex; i++) {
 252             batchRecoveredLogEntry(replicatedLog.get(i));
 253         }
 254
 255         context.setLastApplied(toIndex);
 256         context.setCommitIndex(toIndex);
 257     }
 258
 259     private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
 260         initRecoveryTimer();
 261
 262         int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
 263         if(currentRecoveryBatchCount == 0) {
 264             startLogRecoveryBatch(batchSize);
 265         }
 266
 267         appendRecoveredLogEntry(logEntry.getData());
 268
 269         if(++currentRecoveryBatchCount >= batchSize) {
 270             endCurrentLogRecoveryBatch();
 271         }
 272     }
 273
 274     private void endCurrentLogRecoveryBatch() {
 275         applyCurrentLogRecoveryBatch();
 276         currentRecoveryBatchCount = 0;
 277     }
 278
 279     private void onRecoveryCompletedMessage() {
 280         if(currentRecoveryBatchCount > 0) {
 281             endCurrentLogRecoveryBatch();
 282         }
 283
 284         onRecoveryComplete();
 285
 286         String recoveryTime = "";
 287         if(recoveryTimer != null) {
 288             recoveryTimer.stop();
 289             recoveryTime = " in " + recoveryTimer.toString();
 290             recoveryTimer = null;
 291         }
 292
 293         LOG.info(
 294             "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
 295                 "Persistence Id =  " + persistenceId() +
 296                 " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
 297                 "journal-size={}",
 298             replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(),
 299             replicatedLog.getSnapshotTerm(), replicatedLog.size());
 300
 301         initializeBehavior();
 302     }
 303
 304     protected void initializeBehavior(){
 305         changeCurrentBehavior(new Follower(context));
 306     }
 307
 308     protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
 309         RaftActorBehavior oldBehavior = currentBehavior;
 310         currentBehavior = newBehavior;
 311         handleBehaviorChange(oldBehavior, currentBehavior);
 312     }
 313
 314     @Override public void handleCommand(Object message) {
 315         if (message instanceof ApplyState){
 316             ApplyState applyState = (ApplyState) message;
 317
 318             long elapsedTime = (System.nanoTime() - applyState.getStartTime());
 319             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
 320                 LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
 321                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
 322             }
 323
 324             if(LOG.isDebugEnabled()) {
 325                 LOG.debug("{}: Applying state for log index {} data {}",
 326                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
 327                     applyState.getReplicatedLogEntry().getData());
 328             }
 329
 330             applyState(applyState.getClientActor(), applyState.getIdentifier(),
 331                 applyState.getReplicatedLogEntry().getData());
 332
 333         } else if (message instanceof ApplyJournalEntries){
 334             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
 335             if(LOG.isDebugEnabled()) {
 336                 LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), applyEntries.getToIndex());
 337             }
 338
 339             persistence().persist(applyEntries, APPLY_JOURNAL_ENTRIES_PERSIST_CALLBACK);
 340
 341         } else if(message instanceof ApplySnapshot ) {
 342             Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
 343
 344             if(LOG.isDebugEnabled()) {
 345                 LOG.debug("{}: ApplySnapshot called on Follower Actor " +
 346                         "snapshotIndex:{}, snapshotTerm:{}", persistenceId(), snapshot.getLastAppliedIndex(),
 347                     snapshot.getLastAppliedTerm()
 348                 );
 349             }
 350
 351             applySnapshot(snapshot.getState());
 352
 353             //clears the followers log, sets the snapshot index to ensure adjusted-index works
 354             replicatedLog = new ReplicatedLogImpl(snapshot);
 355             context.setReplicatedLog(replicatedLog);
 356             context.setLastApplied(snapshot.getLastAppliedIndex());
 357
 358         } else if (message instanceof FindLeader) {
 359             getSender().tell(
 360                 new FindLeaderReply(getLeaderAddress()),
 361                 getSelf()
 362             );
 363
 364         } else if (message instanceof SaveSnapshotSuccess) {
 365             SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
 366             LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId());
 367
 368             long sequenceNumber = success.metadata().sequenceNr();
 369
 370             commitSnapshot(sequenceNumber);
 371
 372         } else if (message instanceof SaveSnapshotFailure) {
 373             SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
 374
 375             LOG.error("{}: SaveSnapshotFailure received for snapshot Cause:",
 376                     persistenceId(), saveSnapshotFailure.cause());
 377
 378             context.getReplicatedLog().snapshotRollback();
 379
 380             LOG.info("{}: Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
 381                 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}", persistenceId(),
 382                 context.getReplicatedLog().getSnapshotIndex(),
 383                 context.getReplicatedLog().getSnapshotTerm(),
 384                 context.getReplicatedLog().size());
 385
 386         } else if (message instanceof CaptureSnapshot) {
 387             LOG.debug("{}: CaptureSnapshot received by actor: {}", persistenceId(), message);
 388
 389             if(captureSnapshot == null) {
 390                 captureSnapshot = (CaptureSnapshot)message;
 391                 createSnapshot();
 392             }
 393
 394         } else if (message instanceof CaptureSnapshotReply){
 395             handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot());
 396         } else if(message instanceof GetOnDemandRaftState) {
 397             onGetOnDemandRaftStats();
 398         } else {
 399             RaftActorBehavior oldBehavior = currentBehavior;
 400             currentBehavior = currentBehavior.handleMessage(getSender(), message);
 401
 402             handleBehaviorChange(oldBehavior, currentBehavior);
 403         }
 404     }
 405
 406     private void onGetOnDemandRaftStats() {
 407         // Debugging message to retrieve raft stats.
 408
 409         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
 410                 .commitIndex(context.getCommitIndex())
 411                 .currentTerm(context.getTermInformation().getCurrentTerm())
 412                 .inMemoryJournalDataSize(replicatedLog.dataSize())
 413                 .inMemoryJournalLogSize(replicatedLog.size())
 414                 .isSnapshotCaptureInitiated(context.isSnapshotCaptureInitiated())
 415                 .lastApplied(context.getLastApplied())
 416                 .lastIndex(replicatedLog.lastIndex())
 417                 .lastTerm(replicatedLog.lastTerm())
 418                 .leader(getLeaderId())
 419                 .raftState(currentBehavior.state().toString())
 420                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
 421                 .snapshotIndex(replicatedLog.getSnapshotIndex())
 422                 .snapshotTerm(replicatedLog.getSnapshotTerm())
 423                 .votedFor(context.getTermInformation().getVotedFor())
 424                 .peerAddresses(ImmutableMap.copyOf(context.getPeerAddresses()));
 425
 426         ReplicatedLogEntry lastLogEntry = getLastLogEntry();
 427         if (lastLogEntry != null) {
 428             builder.lastLogIndex(lastLogEntry.getIndex());
 429             builder.lastLogTerm(lastLogEntry.getTerm());
 430         }
 431
 432         if(currentBehavior instanceof AbstractLeader) {
 433             AbstractLeader leader = (AbstractLeader)currentBehavior;
 434             Collection<String> followerIds = leader.getFollowerIds();
 435             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
 436             for(String id: followerIds) {
 437                 final FollowerLogInformation info = leader.getFollower(id);
 438                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
 439                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity())));
 440             }
 441
 442             builder.followerInfoList(followerInfoList);
 443         }
 444
 445         sender().tell(builder.build(), self());
 446
 447     }
 448
 449     private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) {
 450         if (oldBehavior != currentBehavior){
 451             onStateChanged();
 452         }
 453
 454         String oldBehaviorLeaderId = oldBehavior == null? null : oldBehavior.getLeaderId();
 455         String oldBehaviorState = oldBehavior == null? null : oldBehavior.state().name();
 456
 457         // it can happen that the state has not changed but the leader has changed.
 458         onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
 459
 460         if (getRoleChangeNotifier().isPresent() &&
 461                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
 462             getRoleChangeNotifier().get().tell(
 463                     new RoleChanged(getId(), oldBehaviorState , currentBehavior.state().name()),
 464                     getSelf());
 465         }
 466     }
 467
 468     /**
 469      * When a derived RaftActor needs to persist something it must call
 470      * persistData.
 471      *
 472      * @param clientActor
 473      * @param identifier
 474      * @param data
 475      */
 476     protected void persistData(final ActorRef clientActor, final String identifier,
 477         final Payload data) {
 478
 479         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
 480             context.getReplicatedLog().lastIndex() + 1,
 481             context.getTermInformation().getCurrentTerm(), data);
 482
 483         if(LOG.isDebugEnabled()) {
 484             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
 485         }
 486
 487         final RaftActorContext raftContext = getRaftActorContext();
 488
 489         replicatedLog
 490                 .appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
 491                     @Override
 492                     public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
 493                         if(!hasFollowers()){
 494                             // Increment the Commit Index and the Last Applied values
 495                             raftContext.setCommitIndex(replicatedLogEntry.getIndex());
 496                             raftContext.setLastApplied(replicatedLogEntry.getIndex());
 497
 498                             // Apply the state immediately
 499                             applyState(clientActor, identifier, data);
 500
 501                             // Send a ApplyJournalEntries message so that we write the fact that we applied
 502                             // the state to durable storage
 503                             self().tell(new ApplyJournalEntries(replicatedLogEntry.getIndex()), self());
 504
 505                             // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot
 506                             if(!context.isSnapshotCaptureInitiated()){
 507                                 raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(),
 508                                         raftContext.getTermInformation().getCurrentTerm());
 509                                 raftContext.getReplicatedLog().snapshotCommit();
 510                             } else {
 511                                 LOG.debug("{}: Skipping fake snapshotting for {} because real snapshotting is in progress",
 512                                         persistenceId(), getId());
 513                             }
 514                         } else if (clientActor != null) {
 515                             // Send message for replication
 516                             currentBehavior.handleMessage(getSelf(),
 517                                     new Replicate(clientActor, identifier,
 518                                             replicatedLogEntry)
 519                             );
 520                         }
 521
 522                     }
 523                 });    }
 524
 525     protected String getId() {
 526         return context.getId();
 527     }
 528
 529     /**
 530      * Derived actors can call the isLeader method to check if the current
 531      * RaftActor is the Leader or not
 532      *
 533      * @return true it this RaftActor is a Leader false otherwise
 534      */
 535     protected boolean isLeader() {
 536         return context.getId().equals(currentBehavior.getLeaderId());
 537     }
 538
 539     /**
 540      * Derived actor can call getLeader if they need a reference to the Leader.
 541      * This would be useful for example in forwarding a request to an actor
 542      * which is the leader
 543      *
 544      * @return A reference to the leader if known, null otherwise
 545      */
 546     protected ActorSelection getLeader(){
 547         String leaderAddress = getLeaderAddress();
 548
 549         if(leaderAddress == null){
 550             return null;
 551         }
 552
 553         return context.actorSelection(leaderAddress);
 554     }
 555
 556     /**
 557      *
 558      * @return the current leader's id
 559      */
 560     protected String getLeaderId(){
 561         return currentBehavior.getLeaderId();
 562     }
 563
 564     protected RaftState getRaftState() {
 565         return currentBehavior.state();
 566     }
 567
 568     protected ReplicatedLogEntry getLastLogEntry() {
 569         return replicatedLog.last();
 570     }
 571
 572     protected Long getCurrentTerm(){
 573         return context.getTermInformation().getCurrentTerm();
 574     }
 575
 576     protected Long getCommitIndex(){
 577         return context.getCommitIndex();
 578     }
 579
 580     protected Long getLastApplied(){
 581         return context.getLastApplied();
 582     }
 583
 584     protected RaftActorContext getRaftActorContext() {
 585         return context;
 586     }
 587
 588     protected void updateConfigParams(ConfigParams configParams) {
 589         context.setConfigParams(configParams);
 590     }
 591
 592     /**
 593      * setPeerAddress sets the address of a known peer at a later time.
 594      * <p>
 595      * This is to account for situations where a we know that a peer
 596      * exists but we do not know an address up-front. This may also be used in
 597      * situations where a known peer starts off in a different location and we
 598      * need to change it's address
 599      * <p>
 600      * Note that if the peerId does not match the list of peers passed to
 601      * this actor during construction an IllegalStateException will be thrown.
 602      *
 603      * @param peerId
 604      * @param peerAddress
 605      */
 606     protected void setPeerAddress(String peerId, String peerAddress){
 607         context.setPeerAddress(peerId, peerAddress);
 608     }
 609
 610     protected void commitSnapshot(long sequenceNumber) {
 611         context.getReplicatedLog().snapshotCommit();
 612
 613         // TODO: Not sure if we want to be this aggressive with trimming stuff
 614         trimPersistentData(sequenceNumber);
 615     }
 616
 617     /**
 618      * The applyState method will be called by the RaftActor when some data
 619      * needs to be applied to the actor's state
 620      *
 621      * @param clientActor A reference to the client who sent this message. This
 622      *                    is the same reference that was passed to persistData
 623      *                    by the derived actor. clientActor may be null when
 624      *                    the RaftActor is behaving as a follower or during
 625      *                    recovery.
 626      * @param identifier  The identifier of the persisted data. This is also
 627      *                    the same identifier that was passed to persistData by
 628      *                    the derived actor. identifier may be null when
 629      *                    the RaftActor is behaving as a follower or during
 630      *                    recovery
 631      * @param data        A piece of data that was persisted by the persistData call.
 632      *                    This should NEVER be null.
 633      */
 634     protected abstract void applyState(ActorRef clientActor, String identifier,
 635         Object data);
 636
 637     /**
 638      * This method is called during recovery at the start of a batch of state entries. Derived
 639      * classes should perform any initialization needed to start a batch.
 640      */
 641     protected abstract void startLogRecoveryBatch(int maxBatchSize);
 642
 643     /**
 644      * This method is called during recovery to append state data to the current batch. This method
 645      * is called 1 or more times after {@link #startLogRecoveryBatch}.
 646      *
 647      * @param data the state data
 648      */
 649     protected abstract void appendRecoveredLogEntry(Payload data);
 650
 651     /**
 652      * This method is called during recovery to reconstruct the state of the actor.
 653      *
 654      * @param snapshotBytes A snapshot of the state of the actor
 655      */
 656     protected abstract void applyRecoverySnapshot(byte[] snapshotBytes);
 657
 658     /**
 659      * This method is called during recovery at the end of a batch to apply the current batched
 660      * log entries. This method is called after {@link #appendRecoveredLogEntry}.
 661      */
 662     protected abstract void applyCurrentLogRecoveryBatch();
 663
 664     /**
 665      * This method is called when recovery is complete.
 666      */
 667     protected abstract void onRecoveryComplete();
 668
 669     /**
 670      * This method will be called by the RaftActor when a snapshot needs to be
 671      * created. The derived actor should respond with its current state.
 672      * <p/>
 673      * During recovery the state that is returned by the derived actor will
 674      * be passed back to it by calling the applySnapshot  method
 675      *
 676      * @return The current state of the actor
 677      */
 678     protected abstract void createSnapshot();
 679
 680     /**
 681      * This method can be called at any other point during normal
 682      * operations when the derived actor is out of sync with it's peers
 683      * and the only way to bring it in sync is by applying a snapshot
 684      *
 685      * @param snapshotBytes A snapshot of the state of the actor
 686      */
 687     protected abstract void applySnapshot(byte[] snapshotBytes);
 688
 689     /**
 690      * This method will be called by the RaftActor when the state of the
 691      * RaftActor changes. The derived actor can then use methods like
 692      * isLeader or getLeader to do something useful
 693      */
 694     protected abstract void onStateChanged();
 695
 696     protected abstract DataPersistenceProvider persistence();
 697
 698     /**
 699      * Notifier Actor for this RaftActor to notify when a role change happens
 700      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
 701      */
 702     protected abstract Optional<ActorRef> getRoleChangeNotifier();
 703
 704     protected void onLeaderChanged(String oldLeader, String newLeader){};
 705
 706     private void trimPersistentData(long sequenceNumber) {
 707         // Trim akka snapshots
 708         // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
 709         // For now guessing that it is ANDed.
 710         persistence().deleteSnapshots(new SnapshotSelectionCriteria(
 711             sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
 712
 713         // Trim akka journal
 714         persistence().deleteMessages(sequenceNumber);
 715     }
 716
 717     private String getLeaderAddress(){
 718         if(isLeader()){
 719             return getSelf().path().toString();
 720         }
 721         String leaderId = currentBehavior.getLeaderId();
 722         if (leaderId == null) {
 723             return null;
 724         }
 725         String peerAddress = context.getPeerAddress(leaderId);
 726         if(LOG.isDebugEnabled()) {
 727             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
 728                     persistenceId(), leaderId, peerAddress);
 729         }
 730
 731         return peerAddress;
 732     }
 733
 734     private void handleCaptureSnapshotReply(byte[] snapshotBytes) {
 735         LOG.debug("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length);
 736
 737         // create a snapshot object from the state provided and save it
 738         // when snapshot is saved async, SaveSnapshotSuccess is raised.
 739
 740         Snapshot sn = Snapshot.create(snapshotBytes,
 741             context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
 742             captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
 743             captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
 744
 745         persistence().saveSnapshot(sn);
 746
 747         LOG.info("{}: Persisting of snapshot done:{}", persistenceId(), sn.getLogMessage());
 748
 749         long dataThreshold = getTotalMemory() *
 750                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 751         if (context.getReplicatedLog().dataSize() > dataThreshold) {
 752
 753             if(LOG.isDebugEnabled()) {
 754                 LOG.debug("{}: dataSize {} exceeds dataThreshold {} - doing snapshotPreCommit with index {}",
 755                         persistenceId(), context.getReplicatedLog().dataSize(), dataThreshold,
 756                         captureSnapshot.getLastAppliedIndex());
 757             }
 758
 759             // if memory is less, clear the log based on lastApplied.
 760             // this could/should only happen if one of the followers is down
 761             // as normally we keep removing from the log when its replicated to all.
 762             context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getLastAppliedIndex(),
 763                     captureSnapshot.getLastAppliedTerm());
 764
 765             // Don't reset replicatedToAllIndex to -1 as this may prevent us from trimming the log after an
 766             // install snapshot to a follower.
 767             if(captureSnapshot.getReplicatedToAllIndex() >= 0) {
 768                 getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
 769             }
 770         } else if(captureSnapshot.getReplicatedToAllIndex() != -1){
 771             // clear the log based on replicatedToAllIndex
 772             context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getReplicatedToAllIndex(),
 773                     captureSnapshot.getReplicatedToAllTerm());
 774
 775             getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
 776         } else {
 777             // The replicatedToAllIndex was not found in the log
 778             // This means that replicatedToAllIndex never moved beyond -1 or that it is already in the snapshot.
 779             // In this scenario we may need to save the snapshot to the akka persistence
 780             // snapshot for recovery but we do not need to do the replicated log trimming.
 781             context.getReplicatedLog().snapshotPreCommit(replicatedLog.getSnapshotIndex(),
 782                     replicatedLog.getSnapshotTerm());
 783         }
 784
 785
 786         LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex: {} " +
 787             "and term: {}", persistenceId(), replicatedLog.getSnapshotIndex(),
 788             replicatedLog.getSnapshotTerm());
 789
 790         if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
 791             // this would be call straight to the leader and won't initiate in serialization
 792             currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(
 793                     ByteString.copyFrom(snapshotBytes)));
 794         }
 795
 796         captureSnapshot = null;
 797         context.setSnapshotCaptureInitiated(false);
 798     }
 799
 800     protected long getTotalMemory() {
 801         return Runtime.getRuntime().totalMemory();
 802     }
 803
 804     protected boolean hasFollowers(){
 805         return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
 806     }
 807
 808     private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
 809
 810         private static final int DATA_SIZE_DIVIDER = 5;
 811         private long dataSizeSinceLastSnapshot = 0;
 812
 813         public ReplicatedLogImpl(Snapshot snapshot) {
 814             super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
 815                 snapshot.getUnAppliedEntries());
 816         }
 817
 818         public ReplicatedLogImpl() {
 819             super();
 820         }
 821
 822         @Override public void removeFromAndPersist(long logEntryIndex) {
 823             int adjustedIndex = adjustedIndex(logEntryIndex);
 824
 825             if (adjustedIndex < 0) {
 826                 return;
 827             }
 828
 829             // FIXME: Maybe this should be done after the command is saved
 830             journal.subList(adjustedIndex , journal.size()).clear();
 831
 832             persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>() {
 833
 834                 @Override
 835                 public void apply(DeleteEntries param)
 836                         throws Exception {
 837                     //FIXME : Doing nothing for now
 838                     dataSize = 0;
 839                     for (ReplicatedLogEntry entry : journal) {
 840                         dataSize += entry.size();
 841                     }
 842                 }
 843             });
 844         }
 845
 846         @Override public void appendAndPersist(
 847             final ReplicatedLogEntry replicatedLogEntry) {
 848             appendAndPersist(replicatedLogEntry, null);
 849         }
 850
 851         public void appendAndPersist(
 852             final ReplicatedLogEntry replicatedLogEntry,
 853             final Procedure<ReplicatedLogEntry> callback)  {
 854
 855             if(LOG.isDebugEnabled()) {
 856                 LOG.debug("{}: Append log entry and persist {} ", persistenceId(), replicatedLogEntry);
 857             }
 858
 859             // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
 860             journal.add(replicatedLogEntry);
 861
 862             // When persisting events with persist it is guaranteed that the
 863             // persistent actor will not receive further commands between the
 864             // persist call and the execution(s) of the associated event
 865             // handler. This also holds for multiple persist calls in context
 866             // of a single command.
 867             persistence().persist(replicatedLogEntry,
 868                 new Procedure<ReplicatedLogEntry>() {
 869                     @Override
 870                     public void apply(ReplicatedLogEntry evt) throws Exception {
 871                         int logEntrySize = replicatedLogEntry.size();
 872
 873                         dataSize += logEntrySize;
 874                         long dataSizeForCheck = dataSize;
 875
 876                         dataSizeSinceLastSnapshot += logEntrySize;
 877                         long journalSize = lastIndex() + 1;
 878
 879                         if(!hasFollowers()) {
 880                             // When we do not have followers we do not maintain an in-memory log
 881                             // due to this the journalSize will never become anything close to the
 882                             // snapshot batch count. In fact will mostly be 1.
 883                             // Similarly since the journal's dataSize depends on the entries in the
 884                             // journal the journal's dataSize will never reach a value close to the
 885                             // memory threshold.
 886                             // By maintaining the dataSize outside the journal we are tracking essentially
 887                             // what we have written to the disk however since we no longer are in
 888                             // need of doing a snapshot just for the sake of freeing up memory we adjust
 889                             // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
 890                             // as if we were maintaining a real snapshot
 891                             dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
 892                         }
 893
 894                         long dataThreshold = getTotalMemory() *
 895                                 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
 896
 897                         // when a snaphsot is being taken, captureSnapshot != null
 898                         if (!context.isSnapshotCaptureInitiated() &&
 899                                 ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
 900                                         dataSizeForCheck > dataThreshold)) {
 901
 902                             dataSizeSinceLastSnapshot = 0;
 903
 904                             LOG.info("{}: Initiating Snapshot Capture, journalSize = {}, dataSizeForCheck = {}," +
 905                                 " dataThreshold = {}", persistenceId(), journalSize, dataSizeForCheck, dataThreshold);
 906
 907                             long lastAppliedIndex = -1;
 908                             long lastAppliedTerm = -1;
 909
 910                             ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
 911                             if (!hasFollowers()) {
 912                                 lastAppliedIndex = replicatedLogEntry.getIndex();
 913                                 lastAppliedTerm = replicatedLogEntry.getTerm();
 914                             } else if (lastAppliedEntry != null) {
 915                                 lastAppliedIndex = lastAppliedEntry.getIndex();
 916                                 lastAppliedTerm = lastAppliedEntry.getTerm();
 917                             }
 918
 919                             if(LOG.isDebugEnabled()) {
 920                                 LOG.debug("{}: Snapshot Capture logSize: {}", persistenceId(), journal.size());
 921                                 LOG.debug("{}: Snapshot Capture lastApplied:{} ",
 922                                         persistenceId(), context.getLastApplied());
 923                                 LOG.debug("{}: Snapshot Capture lastAppliedIndex:{}", persistenceId(),
 924                                         lastAppliedIndex);
 925                                 LOG.debug("{}: Snapshot Capture lastAppliedTerm:{}", persistenceId(),
 926                                         lastAppliedTerm);
 927                             }
 928
 929                             // send a CaptureSnapshot to self to make the expensive operation async.
 930                             long replicatedToAllIndex = getCurrentBehavior().getReplicatedToAllIndex();
 931                             ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
 932                             getSelf().tell(new CaptureSnapshot(lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
 933                                 (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
 934                                 (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1)),
 935                                 null);
 936                             context.setSnapshotCaptureInitiated(true);
 937                         }
 938                         if (callback != null){
 939                             callback.apply(replicatedLogEntry);
 940                         }
 941                     }
 942                 }
 943             );
 944         }
 945
 946     }
 947
 948     static class DeleteEntries implements Serializable {
 949         private static final long serialVersionUID = 1L;
 950         private final int fromIndex;
 951
 952         public DeleteEntries(int fromIndex) {
 953             this.fromIndex = fromIndex;
 954         }
 955
 956         public int getFromIndex() {
 957             return fromIndex;
 958         }
 959     }
 960
 961
 962     private class ElectionTermImpl implements ElectionTerm {
 963         /**
 964          * Identifier of the actor whose election term information this is
 965          */
 966         private long currentTerm = 0;
 967         private String votedFor = null;
 968
 969         @Override
 970         public long getCurrentTerm() {
 971             return currentTerm;
 972         }
 973
 974         @Override
 975         public String getVotedFor() {
 976             return votedFor;
 977         }
 978
 979         @Override public void update(long currentTerm, String votedFor) {
 980             if(LOG.isDebugEnabled()) {
 981                 LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
 982             }
 983             this.currentTerm = currentTerm;
 984             this.votedFor = votedFor;
 985         }
 986
 987         @Override
 988         public void updateAndPersist(long currentTerm, String votedFor){
 989             update(currentTerm, votedFor);
 990             // FIXME : Maybe first persist then update the state
 991             persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
 992
 993                 @Override public void apply(UpdateElectionTerm param)
 994                     throws Exception {
 995
 996                 }
 997             });
 998         }
 999     }
1000
1001     static class UpdateElectionTerm implements Serializable {
1002         private static final long serialVersionUID = 1L;
1003         private final long currentTerm;
1004         private final String votedFor;
1005
1006         public UpdateElectionTerm(long currentTerm, String votedFor) {
1007             this.currentTerm = currentTerm;
1008             this.votedFor = votedFor;
1009         }
1010
1011         public long getCurrentTerm() {
1012             return currentTerm;
1013         }
1014
1015         public String getVotedFor() {
1016             return votedFor;
1017         }
1018     }
1019
1020     protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
1021
1022         public NonPersistentRaftDataProvider(){
1023
1024         }
1025
1026         /**
1027          * The way snapshotting works is,
1028          * <ol>
1029          * <li> RaftActor calls createSnapshot on the Shard
1030          * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
1031          * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
1032          * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
1033          * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
1034          * in SaveSnapshotSuccess.
1035          * </ol>
1036          * @param o
1037          */
1038         @Override
1039         public void saveSnapshot(Object o) {
1040             // Make saving Snapshot successful
1041             commitSnapshot(-1L);
1042         }
1043     }
1044
1045     @VisibleForTesting
1046     void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
1047         currentBehavior = behavior;
1048     }
1049
1050     protected RaftActorBehavior getCurrentBehavior() {
1051         return currentBehavior;
1052     }
1053
1054 }