opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.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.behaviors;
  10
  11 import akka.actor.ActorRef;
  12 import akka.actor.ActorSelection;
  13 import akka.actor.Cancellable;
  14 import com.google.common.annotations.VisibleForTesting;
  15 import com.google.common.base.Optional;
  16 import com.google.common.base.Preconditions;
  17 import com.google.common.collect.ImmutableMap;
  18 import com.google.common.collect.ImmutableMap.Builder;
  19 import com.google.protobuf.ByteString;
  20 import java.io.IOException;
  21 import java.util.Collection;
  22 import java.util.Collections;
  23 import java.util.HashMap;
  24 import java.util.Iterator;
  25 import java.util.LinkedList;
  26 import java.util.List;
  27 import java.util.Map;
  28 import java.util.Map.Entry;
  29 import java.util.concurrent.TimeUnit;
  30 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
  31 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
  32 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
  33 import org.opendaylight.controller.cluster.raft.FollowerLogInformationImpl;
  34 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  35 import org.opendaylight.controller.cluster.raft.RaftState;
  36 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  37 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
  38 import org.opendaylight.controller.cluster.raft.base.messages.InitiateInstallSnapshot;
  39 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  40 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
  41 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  42 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  43 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  44 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  45 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  46 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  47 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  48 import scala.concurrent.duration.FiniteDuration;
  49
  50 /**
  51  * The behavior of a RaftActor when it is in the Leader state
  52  * <p/>
  53  * Leaders:
  54  * <ul>
  55  * <li> Upon election: send initial empty AppendEntries RPCs
  56  * (heartbeat) to each server; repeat during idle periods to
  57  * prevent election timeouts (§5.2)
  58  * <li> If command received from client: append entry to local log,
  59  * respond after entry applied to state machine (§5.3)
  60  * <li> If last log index ≥ nextIndex for a follower: send
  61  * AppendEntries RPC with log entries starting at nextIndex
  62  * <ul>
  63  * <li> If successful: update nextIndex and matchIndex for
  64  * follower (§5.3)
  65  * <li> If AppendEntries fails because of log inconsistency:
  66  * decrement nextIndex and retry (§5.3)
  67  * </ul>
  68  * <li> If there exists an N such that N > commitIndex, a majority
  69  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  70  * set commitIndex = N (§5.3, §5.4).
  71  */
  72 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
  73
  74     // The index of the first chunk that is sent when installing a snapshot
  75     public static final int FIRST_CHUNK_INDEX = 1;
  76
  77     // The index that the follower should respond with if it needs the install snapshot to be reset
  78     public static final int INVALID_CHUNK_INDEX = -1;
  79
  80     // This would be passed as the hash code of the last chunk when sending the first chunk
  81     public static final int INITIAL_LAST_CHUNK_HASH_CODE = -1;
  82
  83     private final Map<String, FollowerLogInformation> followerToLog;
  84     private final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
  85
  86     private Cancellable heartbeatSchedule = null;
  87
  88     private final Collection<ClientRequestTracker> trackerList = new LinkedList<>();
  89
  90     protected final int minReplicationCount;
  91
  92     protected final int minIsolatedLeaderPeerCount;
  93
  94     private Optional<ByteString> snapshot;
  95
  96     public AbstractLeader(RaftActorContext context) {
  97         super(context);
  98
  99         final Builder<String, FollowerLogInformation> ftlBuilder = ImmutableMap.builder();
 100         for (String followerId : context.getPeerAddresses().keySet()) {
 101             FollowerLogInformation followerLogInformation =
 102                 new FollowerLogInformationImpl(followerId,
 103                     context.getCommitIndex(), -1,
 104                     context.getConfigParams().getElectionTimeOutInterval());
 105
 106             ftlBuilder.put(followerId, followerLogInformation);
 107         }
 108         followerToLog = ftlBuilder.build();
 109
 110         leaderId = context.getId();
 111
 112         LOG.debug("Election:Leader has following peers: {}", getFollowerIds());
 113
 114         minReplicationCount = getMajorityVoteCount(getFollowerIds().size());
 115
 116         // the isolated Leader peer count will be 1 less than the majority vote count.
 117         // this is because the vote count has the self vote counted in it
 118         // for e.g
 119         // 0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 120         // 2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 121         // 4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 122         minIsolatedLeaderPeerCount = minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 123
 124         snapshot = Optional.absent();
 125
 126         // Immediately schedule a heartbeat
 127         // Upon election: send initial empty AppendEntries RPCs
 128         // (heartbeat) to each server; repeat during idle periods to
 129         // prevent election timeouts (§5.2)
 130         scheduleHeartBeat(new FiniteDuration(0, TimeUnit.SECONDS));
 131     }
 132
 133     /**
 134      * Return an immutable collection of follower identifiers.
 135      *
 136      * @return Collection of follower IDs
 137      */
 138     protected final Collection<String> getFollowerIds() {
 139         return followerToLog.keySet();
 140     }
 141
 142     private Optional<ByteString> getSnapshot() {
 143         return snapshot;
 144     }
 145
 146     @VisibleForTesting
 147     void setSnapshot(Optional<ByteString> snapshot) {
 148         this.snapshot = snapshot;
 149     }
 150
 151     @Override
 152     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 153         AppendEntries appendEntries) {
 154
 155         if(LOG.isDebugEnabled()) {
 156             LOG.debug(appendEntries.toString());
 157         }
 158
 159         return this;
 160     }
 161
 162     @Override
 163     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 164         AppendEntriesReply appendEntriesReply) {
 165
 166         if(! appendEntriesReply.isSuccess()) {
 167             if(LOG.isDebugEnabled()) {
 168                 LOG.debug(appendEntriesReply.toString());
 169             }
 170         }
 171
 172         // Update the FollowerLogInformation
 173         String followerId = appendEntriesReply.getFollowerId();
 174         FollowerLogInformation followerLogInformation =
 175             followerToLog.get(followerId);
 176
 177         if(followerLogInformation == null){
 178             LOG.error("Unknown follower {}", followerId);
 179             return this;
 180         }
 181
 182         followerLogInformation.markFollowerActive();
 183
 184         if (appendEntriesReply.isSuccess()) {
 185             followerLogInformation
 186                 .setMatchIndex(appendEntriesReply.getLogLastIndex());
 187             followerLogInformation
 188                 .setNextIndex(appendEntriesReply.getLogLastIndex() + 1);
 189         } else {
 190
 191             // TODO: When we find that the follower is out of sync with the
 192             // Leader we simply decrement that followers next index by 1.
 193             // Would it be possible to do better than this? The RAFT spec
 194             // does not explicitly deal with it but may be something for us to
 195             // think about
 196
 197             followerLogInformation.decrNextIndex();
 198         }
 199
 200         // Now figure out if this reply warrants a change in the commitIndex
 201         // If there exists an N such that N > commitIndex, a majority
 202         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 203         // set commitIndex = N (§5.3, §5.4).
 204         for (long N = context.getCommitIndex() + 1; ; N++) {
 205             int replicatedCount = 1;
 206
 207             for (FollowerLogInformation info : followerToLog.values()) {
 208                 if (info.getMatchIndex() >= N) {
 209                     replicatedCount++;
 210                 }
 211             }
 212
 213             if (replicatedCount >= minReplicationCount) {
 214                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 215                 if (replicatedLogEntry != null &&
 216                     replicatedLogEntry.getTerm() == currentTerm()) {
 217                     context.setCommitIndex(N);
 218                 }
 219             } else {
 220                 break;
 221             }
 222         }
 223
 224         // Apply the change to the state machine
 225         if (context.getCommitIndex() > context.getLastApplied()) {
 226             applyLogToStateMachine(context.getCommitIndex());
 227         }
 228
 229         return this;
 230     }
 231
 232     @Override
 233     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 234         final Iterator<ClientRequestTracker> it = trackerList.iterator();
 235         while (it.hasNext()) {
 236             final ClientRequestTracker t = it.next();
 237             if (t.getIndex() == logIndex) {
 238                 it.remove();
 239                 return t;
 240             }
 241         }
 242
 243         return null;
 244     }
 245
 246     @Override
 247     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 248         for (ClientRequestTracker tracker : trackerList) {
 249             if (tracker.getIndex() == logIndex) {
 250                 return tracker;
 251             }
 252         }
 253         return null;
 254     }
 255
 256     @Override
 257     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 258         RequestVoteReply requestVoteReply) {
 259         return this;
 260     }
 261
 262     @Override
 263     public RaftState state() {
 264         return RaftState.Leader;
 265     }
 266
 267     @Override
 268     public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 269         Preconditions.checkNotNull(sender, "sender should not be null");
 270
 271         Object message = fromSerializableMessage(originalMessage);
 272
 273         if (message instanceof RaftRPC) {
 274             RaftRPC rpc = (RaftRPC) message;
 275             // If RPC request or response contains term T > currentTerm:
 276             // set currentTerm = T, convert to follower (§5.1)
 277             // This applies to all RPC messages and responses
 278             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 279                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 280
 281                 return switchBehavior(new Follower(context));
 282             }
 283         }
 284
 285         try {
 286             if (message instanceof SendHeartBeat) {
 287                 sendHeartBeat();
 288                 return this;
 289
 290             } else if(message instanceof InitiateInstallSnapshot) {
 291                 installSnapshotIfNeeded();
 292
 293             } else if(message instanceof SendInstallSnapshot) {
 294                 // received from RaftActor
 295                 setSnapshot(Optional.of(((SendInstallSnapshot) message).getSnapshot()));
 296                 sendInstallSnapshot();
 297
 298             } else if (message instanceof Replicate) {
 299                 replicate((Replicate) message);
 300
 301             } else if (message instanceof InstallSnapshotReply){
 302                 handleInstallSnapshotReply((InstallSnapshotReply) message);
 303
 304             }
 305         } finally {
 306             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 307         }
 308
 309         return super.handleMessage(sender, message);
 310     }
 311
 312     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 313         String followerId = reply.getFollowerId();
 314         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 315         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 316         followerLogInformation.markFollowerActive();
 317
 318         if (followerToSnapshot != null &&
 319             followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 320
 321             if (reply.isSuccess()) {
 322                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 323                     //this was the last chunk reply
 324                     if(LOG.isDebugEnabled()) {
 325                         LOG.debug("InstallSnapshotReply received, " +
 326                                 "last chunk received, Chunk:{}. Follower:{} Setting nextIndex:{}",
 327                             reply.getChunkIndex(), followerId,
 328                             context.getReplicatedLog().getSnapshotIndex() + 1
 329                         );
 330                     }
 331
 332                     followerLogInformation.setMatchIndex(
 333                         context.getReplicatedLog().getSnapshotIndex());
 334                     followerLogInformation.setNextIndex(
 335                         context.getReplicatedLog().getSnapshotIndex() + 1);
 336                     mapFollowerToSnapshot.remove(followerId);
 337
 338                     if(LOG.isDebugEnabled()) {
 339                         LOG.debug("followerToLog.get(followerId).getNextIndex()=" +
 340                             followerToLog.get(followerId).getNextIndex());
 341                     }
 342
 343                     if (mapFollowerToSnapshot.isEmpty()) {
 344                         // once there are no pending followers receiving snapshots
 345                         // we can remove snapshot from the memory
 346                         setSnapshot(Optional.<ByteString>absent());
 347                     }
 348
 349                 } else {
 350                     followerToSnapshot.markSendStatus(true);
 351                 }
 352             } else {
 353                 LOG.info("InstallSnapshotReply received, " +
 354                         "sending snapshot chunk failed, Will retry, Chunk:{}",
 355                     reply.getChunkIndex()
 356                 );
 357
 358                 followerToSnapshot.markSendStatus(false);
 359             }
 360
 361         } else {
 362             LOG.error("ERROR!!" +
 363                     "FollowerId in InstallSnapshotReply not known to Leader" +
 364                     " or Chunk Index in InstallSnapshotReply not matching {} != {}",
 365                 followerToSnapshot.getChunkIndex(), reply.getChunkIndex()
 366             );
 367
 368             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 369                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 370                 // so that Installing the snapshot can resume from the beginning
 371                 followerToSnapshot.reset();
 372             }
 373         }
 374     }
 375
 376     private void replicate(Replicate replicate) {
 377         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 378
 379         if(LOG.isDebugEnabled()) {
 380             LOG.debug("Replicate message {}", logIndex);
 381         }
 382
 383         // Create a tracker entry we will use this later to notify the
 384         // client actor
 385         trackerList.add(
 386             new ClientRequestTrackerImpl(replicate.getClientActor(),
 387                 replicate.getIdentifier(),
 388                 logIndex)
 389         );
 390
 391         if (followerToLog.isEmpty()) {
 392             context.setCommitIndex(logIndex);
 393             applyLogToStateMachine(logIndex);
 394         } else {
 395             sendAppendEntries();
 396         }
 397     }
 398
 399     private void sendAppendEntries() {
 400         // Send an AppendEntries to all followers
 401         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 402             final String followerId = e.getKey();
 403             ActorSelection followerActor = context.getPeerActorSelection(followerId);
 404
 405             if (followerActor != null) {
 406                 FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 407                 long followerNextIndex = followerLogInformation.getNextIndex();
 408                 boolean isFollowerActive = followerLogInformation.isFollowerActive();
 409
 410                 if (mapFollowerToSnapshot.get(followerId) != null) {
 411                     // if install snapshot is in process , then sent next chunk if possible
 412                     if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 413                         sendSnapshotChunk(followerActor, followerId);
 414                     } else {
 415                         // we send a heartbeat even if we have not received a reply for the last chunk
 416                         sendAppendEntriesToFollower(followerActor, followerNextIndex,
 417                             Collections.<ReplicatedLogEntry>emptyList());
 418                     }
 419
 420                 } else {
 421                     long leaderLastIndex = context.getReplicatedLog().lastIndex();
 422                     long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 423                     final List<ReplicatedLogEntry> entries;
 424
 425                     if (isFollowerActive &&
 426                         context.getReplicatedLog().isPresent(followerNextIndex)) {
 427                         // FIXME : Sending one entry at a time
 428                         entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
 429
 430                     } else if (isFollowerActive && followerNextIndex >= 0 &&
 431                         leaderLastIndex >= followerNextIndex ) {
 432                         // if the followers next index is not present in the leaders log, and
 433                         // if the follower is just not starting and if leader's index is more than followers index
 434                         // then snapshot should be sent
 435
 436                         if(LOG.isDebugEnabled()) {
 437                             LOG.debug("InitiateInstallSnapshot to follower:{}," +
 438                                     "follower-nextIndex:{}, leader-snapshot-index:{},  " +
 439                                     "leader-last-index:{}", followerId,
 440                                 followerNextIndex, leaderSnapShotIndex, leaderLastIndex
 441                             );
 442                         }
 443                         actor().tell(new InitiateInstallSnapshot(), actor());
 444
 445                         // we would want to sent AE as the capture snapshot might take time
 446                         entries =  Collections.<ReplicatedLogEntry>emptyList();
 447
 448                     } else {
 449                         //we send an AppendEntries, even if the follower is inactive
 450                         // in-order to update the followers timestamp, in case it becomes active again
 451                         entries =  Collections.<ReplicatedLogEntry>emptyList();
 452                     }
 453
 454                     sendAppendEntriesToFollower(followerActor, followerNextIndex, entries);
 455
 456                 }
 457             }
 458         }
 459     }
 460
 461     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 462         List<ReplicatedLogEntry> entries) {
 463         followerActor.tell(
 464             new AppendEntries(currentTerm(), context.getId(),
 465                 prevLogIndex(followerNextIndex),
 466                 prevLogTerm(followerNextIndex), entries,
 467                 context.getCommitIndex()).toSerializable(),
 468             actor()
 469         );
 470     }
 471
 472     /**
 473      * An installSnapshot is scheduled at a interval that is a multiple of
 474      * a HEARTBEAT_INTERVAL. This is to avoid the need to check for installing
 475      * snapshots at every heartbeat.
 476      *
 477      * Install Snapshot works as follows
 478      * 1. Leader sends a InitiateInstallSnapshot message to self
 479      * 2. Leader then initiates the capture snapshot by sending a CaptureSnapshot message to actor
 480      * 3. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 481      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 482      * 4. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 483      * 5. On complete, Follower sends back a InstallSnapshotReply.
 484      * 6. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 485      * and replenishes the memory by deleting the snapshot in Replicated log.
 486      *
 487      */
 488     private void installSnapshotIfNeeded() {
 489         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 490             final ActorSelection followerActor = context.getPeerActorSelection(e.getKey());
 491
 492             if (followerActor != null) {
 493                 long nextIndex = e.getValue().getNextIndex();
 494
 495                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 496                     context.getReplicatedLog().isInSnapshot(nextIndex)) {
 497                     LOG.info("{} follower needs a snapshot install", e.getKey());
 498                     if (snapshot.isPresent()) {
 499                         // if a snapshot is present in the memory, most likely another install is in progress
 500                         // no need to capture snapshot
 501                         sendSnapshotChunk(followerActor, e.getKey());
 502
 503                     } else {
 504                         initiateCaptureSnapshot();
 505                         //we just need 1 follower who would need snapshot to be installed.
 506                         // when we have the snapshot captured, we would again check (in SendInstallSnapshot)
 507                         // who needs an install and send to all who need
 508                         break;
 509                     }
 510
 511                 }
 512             }
 513         }
 514     }
 515
 516     // on every install snapshot, we try to capture the snapshot.
 517     // Once a capture is going on, another one issued will get ignored by RaftActor.
 518     private void initiateCaptureSnapshot() {
 519         LOG.info("Initiating Snapshot Capture to Install Snapshot, Leader:{}", getLeaderId());
 520         ReplicatedLogEntry lastAppliedEntry = context.getReplicatedLog().get(context.getLastApplied());
 521         long lastAppliedIndex = -1;
 522         long lastAppliedTerm = -1;
 523
 524         if (lastAppliedEntry != null) {
 525             lastAppliedIndex = lastAppliedEntry.getIndex();
 526             lastAppliedTerm = lastAppliedEntry.getTerm();
 527         } else if (context.getReplicatedLog().getSnapshotIndex() > -1)  {
 528             lastAppliedIndex = context.getReplicatedLog().getSnapshotIndex();
 529             lastAppliedTerm = context.getReplicatedLog().getSnapshotTerm();
 530         }
 531
 532         boolean isInstallSnapshotInitiated = true;
 533         actor().tell(new CaptureSnapshot(lastIndex(), lastTerm(),
 534                 lastAppliedIndex, lastAppliedTerm, isInstallSnapshotInitiated),
 535             actor());
 536     }
 537
 538
 539     private void sendInstallSnapshot() {
 540         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 541             ActorSelection followerActor = context.getPeerActorSelection(e.getKey());
 542
 543             if (followerActor != null) {
 544                 long nextIndex = e.getValue().getNextIndex();
 545
 546                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 547                     context.getReplicatedLog().isInSnapshot(nextIndex)) {
 548                     sendSnapshotChunk(followerActor, e.getKey());
 549                 }
 550             }
 551         }
 552     }
 553
 554     /**
 555      *  Sends a snapshot chunk to a given follower
 556      *  InstallSnapshot should qualify as a heartbeat too.
 557      */
 558     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 559         try {
 560             if (snapshot.isPresent()) {
 561                 followerActor.tell(
 562                     new InstallSnapshot(currentTerm(), context.getId(),
 563                         context.getReplicatedLog().getSnapshotIndex(),
 564                         context.getReplicatedLog().getSnapshotTerm(),
 565                         getNextSnapshotChunk(followerId,snapshot.get()),
 566                         mapFollowerToSnapshot.get(followerId).incrementChunkIndex(),
 567                         mapFollowerToSnapshot.get(followerId).getTotalChunks(),
 568                         Optional.of(mapFollowerToSnapshot.get(followerId).getLastChunkHashCode())
 569                     ).toSerializable(),
 570                     actor()
 571                 );
 572                 LOG.info("InstallSnapshot sent to follower {}, Chunk: {}/{}",
 573                     followerActor.path(), mapFollowerToSnapshot.get(followerId).getChunkIndex(),
 574                     mapFollowerToSnapshot.get(followerId).getTotalChunks());
 575             }
 576         } catch (IOException e) {
 577             LOG.error(e, "InstallSnapshot failed for Leader.");
 578         }
 579     }
 580
 581     /**
 582      * Acccepts snaphot as ByteString, enters into map for future chunks
 583      * creates and return a ByteString chunk
 584      */
 585     private ByteString getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 586         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 587         if (followerToSnapshot == null) {
 588             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 589             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 590         }
 591         ByteString nextChunk = followerToSnapshot.getNextChunk();
 592         if (LOG.isDebugEnabled()) {
 593             LOG.debug("Leader's snapshot nextChunk size:{}", nextChunk.size());
 594         }
 595         return nextChunk;
 596     }
 597
 598     private void sendHeartBeat() {
 599         if (!followerToLog.isEmpty()) {
 600             sendAppendEntries();
 601         }
 602     }
 603
 604     private void stopHeartBeat() {
 605         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 606             heartbeatSchedule.cancel();
 607         }
 608     }
 609
 610     private void scheduleHeartBeat(FiniteDuration interval) {
 611         if (followerToLog.isEmpty()) {
 612             // Optimization - do not bother scheduling a heartbeat as there are
 613             // no followers
 614             return;
 615         }
 616
 617         stopHeartBeat();
 618
 619         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 620         // message is sent to itself.
 621         // Scheduling the heartbeat only once here because heartbeats do not
 622         // need to be sent if there are other messages being sent to the remote
 623         // actor.
 624         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 625             interval, context.getActor(), new SendHeartBeat(),
 626             context.getActorSystem().dispatcher(), context.getActor());
 627     }
 628
 629     @Override
 630     public void close() throws Exception {
 631         stopHeartBeat();
 632     }
 633
 634     @Override
 635     public String getLeaderId() {
 636         return context.getId();
 637     }
 638
 639     protected boolean isLeaderIsolated() {
 640         int minPresent = minIsolatedLeaderPeerCount;
 641         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 642             if (followerLogInformation.isFollowerActive()) {
 643                 --minPresent;
 644                 if (minPresent == 0) {
 645                     break;
 646                 }
 647             }
 648         }
 649         return (minPresent != 0);
 650     }
 651
 652     /**
 653      * Encapsulates the snapshot bytestring and handles the logic of sending
 654      * snapshot chunks
 655      */
 656     protected class FollowerToSnapshot {
 657         private ByteString snapshotBytes;
 658         private int offset = 0;
 659         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 660         private int replyReceivedForOffset;
 661         // if replyStatus is false, the previous chunk is attempted
 662         private boolean replyStatus = false;
 663         private int chunkIndex;
 664         private int totalChunks;
 665         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 666         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 667
 668         public FollowerToSnapshot(ByteString snapshotBytes) {
 669             this.snapshotBytes = snapshotBytes;
 670             int size = snapshotBytes.size();
 671             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 672                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 673             if(LOG.isDebugEnabled()) {
 674                 LOG.debug("Snapshot {} bytes, total chunks to send:{}",
 675                     size, totalChunks);
 676             }
 677             replyReceivedForOffset = -1;
 678             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 679         }
 680
 681         public ByteString getSnapshotBytes() {
 682             return snapshotBytes;
 683         }
 684
 685         public int incrementOffset() {
 686             if(replyStatus) {
 687                 // if prev chunk failed, we would want to sent the same chunk again
 688                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 689             }
 690             return offset;
 691         }
 692
 693         public int incrementChunkIndex() {
 694             if (replyStatus) {
 695                 // if prev chunk failed, we would want to sent the same chunk again
 696                 chunkIndex =  chunkIndex + 1;
 697             }
 698             return chunkIndex;
 699         }
 700
 701         public int getChunkIndex() {
 702             return chunkIndex;
 703         }
 704
 705         public int getTotalChunks() {
 706             return totalChunks;
 707         }
 708
 709         public boolean canSendNextChunk() {
 710             // we only send a false if a chunk is sent but we have not received a reply yet
 711             return replyReceivedForOffset == offset;
 712         }
 713
 714         public boolean isLastChunk(int chunkIndex) {
 715             return totalChunks == chunkIndex;
 716         }
 717
 718         public void markSendStatus(boolean success) {
 719             if (success) {
 720                 // if the chunk sent was successful
 721                 replyReceivedForOffset = offset;
 722                 replyStatus = true;
 723                 lastChunkHashCode = nextChunkHashCode;
 724             } else {
 725                 // if the chunk sent was failure
 726                 replyReceivedForOffset = offset;
 727                 replyStatus = false;
 728             }
 729         }
 730
 731         public ByteString getNextChunk() {
 732             int snapshotLength = getSnapshotBytes().size();
 733             int start = incrementOffset();
 734             int size = context.getConfigParams().getSnapshotChunkSize();
 735             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 736                 size = snapshotLength;
 737             } else {
 738                 if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 739                     size = snapshotLength - start;
 740                 }
 741             }
 742
 743             if(LOG.isDebugEnabled()) {
 744                 LOG.debug("length={}, offset={},size={}",
 745                     snapshotLength, start, size);
 746             }
 747             ByteString substring = getSnapshotBytes().substring(start, start + size);
 748             nextChunkHashCode = substring.hashCode();
 749             return substring;
 750         }
 751
 752         /**
 753          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 754          */
 755         public void reset(){
 756             offset = 0;
 757             replyStatus = false;
 758             replyReceivedForOffset = offset;
 759             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 760             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 761         }
 762
 763         public int getLastChunkHashCode() {
 764             return lastChunkHashCode;
 765         }
 766     }
 767
 768     // called from example-actor for printing the follower-states
 769     public String printFollowerStates() {
 770         final StringBuilder sb = new StringBuilder();
 771
 772         sb.append('[');
 773         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 774             sb.append('{');
 775             sb.append(followerLogInformation.getId());
 776             sb.append(" state:");
 777             sb.append(followerLogInformation.isFollowerActive());
 778             sb.append("},");
 779         }
 780         sb.append(']');
 781
 782         return sb.toString();
 783     }
 784
 785     @VisibleForTesting
 786     public FollowerLogInformation getFollower(String followerId) {
 787         return followerToLog.get(followerId);
 788     }
 789
 790     @VisibleForTesting
 791     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 792         mapFollowerToSnapshot.put(followerId, snapshot);
 793     }
 794
 795     @VisibleForTesting
 796     public int followerSnapshotSize() {
 797         return mapFollowerToSnapshot.size();
 798     }
 799
 800     @VisibleForTesting
 801     public int followerLogSize() {
 802         return followerToLog.size();
 803     }
 804 }