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     private long replicatedToAllIndex = -1;
  97
  98     public AbstractLeader(RaftActorContext context) {
  99         super(context);
 100
 101         final Builder<String, FollowerLogInformation> ftlBuilder = ImmutableMap.builder();
 102         for (String followerId : context.getPeerAddresses().keySet()) {
 103             FollowerLogInformation followerLogInformation =
 104                 new FollowerLogInformationImpl(followerId,
 105                     context.getCommitIndex(), -1,
 106                     context.getConfigParams().getElectionTimeOutInterval());
 107
 108             ftlBuilder.put(followerId, followerLogInformation);
 109         }
 110         followerToLog = ftlBuilder.build();
 111
 112         leaderId = context.getId();
 113
 114         LOG.debug("{}: Election: Leader has following peers: {}", context.getId(), getFollowerIds());
 115
 116         minReplicationCount = getMajorityVoteCount(getFollowerIds().size());
 117
 118         // the isolated Leader peer count will be 1 less than the majority vote count.
 119         // this is because the vote count has the self vote counted in it
 120         // for e.g
 121         // 0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 122         // 2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 123         // 4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 124         minIsolatedLeaderPeerCount = minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 125
 126         snapshot = Optional.absent();
 127
 128         // Immediately schedule a heartbeat
 129         // Upon election: send initial empty AppendEntries RPCs
 130         // (heartbeat) to each server; repeat during idle periods to
 131         // prevent election timeouts (§5.2)
 132         scheduleHeartBeat(new FiniteDuration(0, TimeUnit.SECONDS));
 133     }
 134
 135     /**
 136      * Return an immutable collection of follower identifiers.
 137      *
 138      * @return Collection of follower IDs
 139      */
 140     protected final Collection<String> getFollowerIds() {
 141         return followerToLog.keySet();
 142     }
 143
 144     private Optional<ByteString> getSnapshot() {
 145         return snapshot;
 146     }
 147
 148     @VisibleForTesting
 149     void setSnapshot(Optional<ByteString> snapshot) {
 150         this.snapshot = snapshot;
 151     }
 152
 153     @Override
 154     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 155         AppendEntries appendEntries) {
 156
 157         if(LOG.isDebugEnabled()) {
 158             LOG.debug("{}: handleAppendEntries: {}", context.getId(), appendEntries);
 159         }
 160
 161         return this;
 162     }
 163
 164     @Override
 165     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 166         AppendEntriesReply appendEntriesReply) {
 167
 168         if(! appendEntriesReply.isSuccess()) {
 169             if(LOG.isDebugEnabled()) {
 170                 LOG.debug("{}: handleAppendEntriesReply: {}", context.getId(), appendEntriesReply);
 171             }
 172         }
 173
 174         // Update the FollowerLogInformation
 175         String followerId = appendEntriesReply.getFollowerId();
 176         FollowerLogInformation followerLogInformation =
 177             followerToLog.get(followerId);
 178
 179         if(followerLogInformation == null){
 180             LOG.error("{}: handleAppendEntriesReply - unknown follower {}", context.getId(), followerId);
 181             return this;
 182         }
 183
 184         followerLogInformation.markFollowerActive();
 185
 186         if (appendEntriesReply.isSuccess()) {
 187             followerLogInformation
 188                 .setMatchIndex(appendEntriesReply.getLogLastIndex());
 189             followerLogInformation
 190                 .setNextIndex(appendEntriesReply.getLogLastIndex() + 1);
 191         } else {
 192
 193             // TODO: When we find that the follower is out of sync with the
 194             // Leader we simply decrement that followers next index by 1.
 195             // Would it be possible to do better than this? The RAFT spec
 196             // does not explicitly deal with it but may be something for us to
 197             // think about
 198
 199             followerLogInformation.decrNextIndex();
 200         }
 201
 202         // Now figure out if this reply warrants a change in the commitIndex
 203         // If there exists an N such that N > commitIndex, a majority
 204         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 205         // set commitIndex = N (§5.3, §5.4).
 206         for (long N = context.getCommitIndex() + 1; ; N++) {
 207             int replicatedCount = 1;
 208
 209             for (FollowerLogInformation info : followerToLog.values()) {
 210                 if (info.getMatchIndex() >= N) {
 211                     replicatedCount++;
 212                 }
 213             }
 214
 215             if (replicatedCount >= minReplicationCount) {
 216                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 217                 if (replicatedLogEntry != null &&
 218                     replicatedLogEntry.getTerm() == currentTerm()) {
 219                     context.setCommitIndex(N);
 220                 }
 221             } else {
 222                 break;
 223             }
 224         }
 225
 226         // Apply the change to the state machine
 227         if (context.getCommitIndex() > context.getLastApplied()) {
 228             applyLogToStateMachine(context.getCommitIndex());
 229         }
 230
 231         if (!context.isSnapshotCaptureInitiated()) {
 232             purgeInMemoryLog();
 233         }
 234
 235         //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
 236         sendUpdatesToFollower(followerId, followerLogInformation, false);
 237         return this;
 238     }
 239
 240     private void purgeInMemoryLog() {
 241         //find the lowest index across followers which has been replicated to all. -1 if there are no followers.
 242         // we would delete the in-mem log from that index on, in-order to minimize mem usage
 243         // we would also share this info thru AE with the followers so that they can delete their log entries as well.
 244         long minReplicatedToAllIndex = followerToLog.isEmpty() ? -1 : Long.MAX_VALUE;
 245         for (FollowerLogInformation info : followerToLog.values()) {
 246             minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
 247         }
 248
 249         replicatedToAllIndex = fakeSnapshot(minReplicatedToAllIndex, replicatedToAllIndex);
 250     }
 251
 252     @Override
 253     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 254         final Iterator<ClientRequestTracker> it = trackerList.iterator();
 255         while (it.hasNext()) {
 256             final ClientRequestTracker t = it.next();
 257             if (t.getIndex() == logIndex) {
 258                 it.remove();
 259                 return t;
 260             }
 261         }
 262
 263         return null;
 264     }
 265
 266     @Override
 267     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 268         for (ClientRequestTracker tracker : trackerList) {
 269             if (tracker.getIndex() == logIndex) {
 270                 return tracker;
 271             }
 272         }
 273         return null;
 274     }
 275
 276     @Override
 277     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 278         RequestVoteReply requestVoteReply) {
 279         return this;
 280     }
 281
 282     @Override
 283     public RaftState state() {
 284         return RaftState.Leader;
 285     }
 286
 287     @Override
 288     public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 289         Preconditions.checkNotNull(sender, "sender should not be null");
 290
 291         Object message = fromSerializableMessage(originalMessage);
 292
 293         if (message instanceof RaftRPC) {
 294             RaftRPC rpc = (RaftRPC) message;
 295             // If RPC request or response contains term T > currentTerm:
 296             // set currentTerm = T, convert to follower (§5.1)
 297             // This applies to all RPC messages and responses
 298             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 299                 LOG.debug("{}: Term {} in \"{}\" message is greater than leader's term {}", context.getId(),
 300                         rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 301
 302                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 303
 304                 return switchBehavior(new Follower(context));
 305             }
 306         }
 307
 308         try {
 309             if (message instanceof SendHeartBeat) {
 310                 sendHeartBeat();
 311                 return this;
 312
 313             } else if(message instanceof InitiateInstallSnapshot) {
 314                 installSnapshotIfNeeded();
 315
 316             } else if(message instanceof SendInstallSnapshot) {
 317                 // received from RaftActor
 318                 setSnapshot(Optional.of(((SendInstallSnapshot) message).getSnapshot()));
 319                 sendInstallSnapshot();
 320
 321             } else if (message instanceof Replicate) {
 322                 replicate((Replicate) message);
 323
 324             } else if (message instanceof InstallSnapshotReply){
 325                 handleInstallSnapshotReply((InstallSnapshotReply) message);
 326
 327             }
 328         } finally {
 329             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 330         }
 331
 332         return super.handleMessage(sender, message);
 333     }
 334
 335     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 336         String followerId = reply.getFollowerId();
 337         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 338
 339         if (followerToSnapshot == null) {
 340             LOG.error("{}: FollowerId {} in InstallSnapshotReply not known to Leader",
 341                     context.getId(), followerId);
 342             return;
 343         }
 344
 345         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 346         followerLogInformation.markFollowerActive();
 347
 348         if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 349             if (reply.isSuccess()) {
 350                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 351                     //this was the last chunk reply
 352                     if(LOG.isDebugEnabled()) {
 353                         LOG.debug("{}: InstallSnapshotReply received, " +
 354                                 "last chunk received, Chunk:{}. Follower:{} Setting nextIndex:{}",
 355                                 context.getId(), reply.getChunkIndex(), followerId,
 356                             context.getReplicatedLog().getSnapshotIndex() + 1
 357                         );
 358                     }
 359
 360                     followerLogInformation.setMatchIndex(
 361                         context.getReplicatedLog().getSnapshotIndex());
 362                     followerLogInformation.setNextIndex(
 363                         context.getReplicatedLog().getSnapshotIndex() + 1);
 364                     mapFollowerToSnapshot.remove(followerId);
 365
 366                     if(LOG.isDebugEnabled()) {
 367                         LOG.debug("{}: followerToLog.get(followerId).getNextIndex()=" +
 368                                 context.getId(), followerToLog.get(followerId).getNextIndex());
 369                     }
 370
 371                     if (mapFollowerToSnapshot.isEmpty()) {
 372                         // once there are no pending followers receiving snapshots
 373                         // we can remove snapshot from the memory
 374                         setSnapshot(Optional.<ByteString>absent());
 375                     }
 376
 377                 } else {
 378                     followerToSnapshot.markSendStatus(true);
 379                 }
 380             } else {
 381                 LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
 382                         context.getId(), reply.getChunkIndex());
 383
 384                 followerToSnapshot.markSendStatus(false);
 385             }
 386         } else {
 387             LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
 388                     context.getId(), reply.getChunkIndex(), followerId,
 389                     followerToSnapshot.getChunkIndex());
 390
 391             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 392                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 393                 // so that Installing the snapshot can resume from the beginning
 394                 followerToSnapshot.reset();
 395             }
 396         }
 397     }
 398
 399     private void replicate(Replicate replicate) {
 400         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 401
 402         if(LOG.isDebugEnabled()) {
 403             LOG.debug("{}: Replicate message {}", context.getId(), logIndex);
 404         }
 405
 406         // Create a tracker entry we will use this later to notify the
 407         // client actor
 408         trackerList.add(
 409             new ClientRequestTrackerImpl(replicate.getClientActor(),
 410                 replicate.getIdentifier(),
 411                 logIndex)
 412         );
 413
 414         if (followerToLog.isEmpty()) {
 415             context.setCommitIndex(logIndex);
 416             applyLogToStateMachine(logIndex);
 417         } else {
 418             sendAppendEntries();
 419         }
 420     }
 421
 422     private void sendAppendEntries() {
 423         // Send an AppendEntries to all followers
 424         long heartbeatInterval = context.getConfigParams().getHeartBeatInterval().toMillis();
 425         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 426             final String followerId = e.getKey();
 427             final FollowerLogInformation followerLogInformation = e.getValue();
 428             // This checks helps not to send a repeat message to the follower
 429             if(followerLogInformation.timeSinceLastActivity() >= heartbeatInterval) {
 430                 sendUpdatesToFollower(followerId, followerLogInformation, true);
 431             }
 432         }
 433     }
 434
 435     /**
 436      *
 437      * This method checks if any update needs to be sent to the given follower. This includes append log entries,
 438      * sending next snapshot chunk, and initiating a snapshot.
 439      * @return true if any update is sent, false otherwise
 440      */
 441
 442     private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
 443                                           boolean sendHeartbeat) {
 444
 445         ActorSelection followerActor = context.getPeerActorSelection(followerId);
 446         if (followerActor != null) {
 447             long followerNextIndex = followerLogInformation.getNextIndex();
 448             boolean isFollowerActive = followerLogInformation.isFollowerActive();
 449
 450             if (mapFollowerToSnapshot.get(followerId) != null) {
 451                 // if install snapshot is in process , then sent next chunk if possible
 452                 if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 453                     sendSnapshotChunk(followerActor, followerId);
 454                 } else if(sendHeartbeat) {
 455                     // we send a heartbeat even if we have not received a reply for the last chunk
 456                     sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
 457                         Collections.<ReplicatedLogEntry>emptyList(), followerId);
 458                 }
 459             } else {
 460                 long leaderLastIndex = context.getReplicatedLog().lastIndex();
 461                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 462                 if (isFollowerActive &&
 463                     context.getReplicatedLog().isPresent(followerNextIndex)) {
 464                     // FIXME : Sending one entry at a time
 465                     final List<ReplicatedLogEntry> entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
 466
 467                     sendAppendEntriesToFollower(followerActor, followerNextIndex, entries, followerId);
 468
 469                 } else if (isFollowerActive && followerNextIndex >= 0 &&
 470                     leaderLastIndex >= followerNextIndex) {
 471                     // if the followers next index is not present in the leaders log, and
 472                     // if the follower is just not starting and if leader's index is more than followers index
 473                     // then snapshot should be sent
 474
 475                     if (LOG.isDebugEnabled()) {
 476                         LOG.debug("InitiateInstallSnapshot to follower:{}," +
 477                                 "follower-nextIndex:{}, leader-snapshot-index:{},  " +
 478                                 "leader-last-index:{}", followerId,
 479                             followerNextIndex, leaderSnapShotIndex, leaderLastIndex
 480                         );
 481                     }
 482                     actor().tell(new InitiateInstallSnapshot(), actor());
 483
 484                     // Send heartbeat to follower whenever install snapshot is initiated.
 485                     sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
 486                             Collections.<ReplicatedLogEntry>emptyList(), followerId);
 487
 488                 } else if(sendHeartbeat) {
 489                     //we send an AppendEntries, even if the follower is inactive
 490                     // in-order to update the followers timestamp, in case it becomes active again
 491                     sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
 492                         Collections.<ReplicatedLogEntry>emptyList(), followerId);
 493                 }
 494
 495             }
 496         }
 497     }
 498
 499     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 500         List<ReplicatedLogEntry> entries, String followerId) {
 501         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
 502             prevLogIndex(followerNextIndex),
 503             prevLogTerm(followerNextIndex), entries,
 504             context.getCommitIndex(), replicatedToAllIndex);
 505
 506         if(!entries.isEmpty()) {
 507             LOG.debug("{}: Sending AppendEntries to follower {}: {}", context.getId(), followerId,
 508                     appendEntries);
 509         }
 510
 511         followerActor.tell(appendEntries.toSerializable(), actor());
 512     }
 513
 514     /**
 515      * An installSnapshot is scheduled at a interval that is a multiple of
 516      * a HEARTBEAT_INTERVAL. This is to avoid the need to check for installing
 517      * snapshots at every heartbeat.
 518      *
 519      * Install Snapshot works as follows
 520      * 1. Leader sends a InitiateInstallSnapshot message to self
 521      * 2. Leader then initiates the capture snapshot by sending a CaptureSnapshot message to actor
 522      * 3. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 523      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 524      * 4. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 525      * 5. On complete, Follower sends back a InstallSnapshotReply.
 526      * 6. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 527      * and replenishes the memory by deleting the snapshot in Replicated log.
 528      *
 529      */
 530     private void installSnapshotIfNeeded() {
 531         if(LOG.isDebugEnabled()) {
 532             LOG.debug("{}: installSnapshotIfNeeded, followers {}", context.getId(), followerToLog.keySet());
 533         }
 534
 535         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 536             final ActorSelection followerActor = context.getPeerActorSelection(e.getKey());
 537
 538             if (followerActor != null) {
 539                 long nextIndex = e.getValue().getNextIndex();
 540
 541                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 542                         context.getReplicatedLog().isInSnapshot(nextIndex)) {
 543                     LOG.info("{}: {} follower needs a snapshot install", context.getId(), e.getKey());
 544                     if (snapshot.isPresent()) {
 545                         // if a snapshot is present in the memory, most likely another install is in progress
 546                         // no need to capture snapshot
 547                         sendSnapshotChunk(followerActor, e.getKey());
 548
 549                     } else {
 550                         initiateCaptureSnapshot();
 551                         //we just need 1 follower who would need snapshot to be installed.
 552                         // when we have the snapshot captured, we would again check (in SendInstallSnapshot)
 553                         // who needs an install and send to all who need
 554                         break;
 555                     }
 556
 557                 }
 558             }
 559         }
 560     }
 561
 562     // on every install snapshot, we try to capture the snapshot.
 563     // Once a capture is going on, another one issued will get ignored by RaftActor.
 564     private void initiateCaptureSnapshot() {
 565         LOG.info("{}: Initiating Snapshot Capture to Install Snapshot, Leader:{}", context.getId(), getLeaderId());
 566         ReplicatedLogEntry lastAppliedEntry = context.getReplicatedLog().get(context.getLastApplied());
 567         long lastAppliedIndex = -1;
 568         long lastAppliedTerm = -1;
 569
 570         if (lastAppliedEntry != null) {
 571             lastAppliedIndex = lastAppliedEntry.getIndex();
 572             lastAppliedTerm = lastAppliedEntry.getTerm();
 573         } else if (context.getReplicatedLog().getSnapshotIndex() > -1)  {
 574             lastAppliedIndex = context.getReplicatedLog().getSnapshotIndex();
 575             lastAppliedTerm = context.getReplicatedLog().getSnapshotTerm();
 576         }
 577
 578         boolean isInstallSnapshotInitiated = true;
 579         actor().tell(new CaptureSnapshot(lastIndex(), lastTerm(),
 580                 lastAppliedIndex, lastAppliedTerm, isInstallSnapshotInitiated),
 581             actor());
 582     }
 583
 584
 585     private void sendInstallSnapshot() {
 586         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 587             ActorSelection followerActor = context.getPeerActorSelection(e.getKey());
 588
 589             if (followerActor != null) {
 590                 long nextIndex = e.getValue().getNextIndex();
 591
 592                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 593                     context.getReplicatedLog().isInSnapshot(nextIndex)) {
 594                     sendSnapshotChunk(followerActor, e.getKey());
 595                 }
 596             }
 597         }
 598     }
 599
 600     /**
 601      *  Sends a snapshot chunk to a given follower
 602      *  InstallSnapshot should qualify as a heartbeat too.
 603      */
 604     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 605         try {
 606             if (snapshot.isPresent()) {
 607                 ByteString nextSnapshotChunk = getNextSnapshotChunk(followerId,snapshot.get());
 608
 609                 // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
 610                 // followerId to the followerToSnapshot map.
 611                 FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 612
 613                 followerActor.tell(
 614                     new InstallSnapshot(currentTerm(), context.getId(),
 615                         context.getReplicatedLog().getSnapshotIndex(),
 616                         context.getReplicatedLog().getSnapshotTerm(),
 617                         nextSnapshotChunk,
 618                         followerToSnapshot.incrementChunkIndex(),
 619                         followerToSnapshot.getTotalChunks(),
 620                         Optional.of(followerToSnapshot.getLastChunkHashCode())
 621                     ).toSerializable(),
 622                     actor()
 623                 );
 624                 LOG.info("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
 625                         context.getId(), followerActor.path(),
 626                         followerToSnapshot.getChunkIndex(),
 627                         followerToSnapshot.getTotalChunks());
 628             }
 629         } catch (IOException e) {
 630             LOG.error(e, "{}: InstallSnapshot failed for Leader.", context.getId());
 631         }
 632     }
 633
 634     /**
 635      * Acccepts snaphot as ByteString, enters into map for future chunks
 636      * creates and return a ByteString chunk
 637      */
 638     private ByteString getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 639         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 640         if (followerToSnapshot == null) {
 641             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 642             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 643         }
 644         ByteString nextChunk = followerToSnapshot.getNextChunk();
 645         if (LOG.isDebugEnabled()) {
 646             LOG.debug("{}: Leader's snapshot nextChunk size:{}", context.getId(), nextChunk.size());
 647         }
 648         return nextChunk;
 649     }
 650
 651     private void sendHeartBeat() {
 652         if (!followerToLog.isEmpty()) {
 653             sendAppendEntries();
 654         }
 655     }
 656
 657     private void stopHeartBeat() {
 658         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 659             heartbeatSchedule.cancel();
 660         }
 661     }
 662
 663     private void scheduleHeartBeat(FiniteDuration interval) {
 664         if (followerToLog.isEmpty()) {
 665             // Optimization - do not bother scheduling a heartbeat as there are
 666             // no followers
 667             return;
 668         }
 669
 670         stopHeartBeat();
 671
 672         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 673         // message is sent to itself.
 674         // Scheduling the heartbeat only once here because heartbeats do not
 675         // need to be sent if there are other messages being sent to the remote
 676         // actor.
 677         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 678             interval, context.getActor(), new SendHeartBeat(),
 679             context.getActorSystem().dispatcher(), context.getActor());
 680     }
 681
 682     @Override
 683     public void close() throws Exception {
 684         stopHeartBeat();
 685     }
 686
 687     @Override
 688     public String getLeaderId() {
 689         return context.getId();
 690     }
 691
 692     protected boolean isLeaderIsolated() {
 693         int minPresent = minIsolatedLeaderPeerCount;
 694         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 695             if (followerLogInformation.isFollowerActive()) {
 696                 --minPresent;
 697                 if (minPresent == 0) {
 698                     break;
 699                 }
 700             }
 701         }
 702         return (minPresent != 0);
 703     }
 704
 705     /**
 706      * Encapsulates the snapshot bytestring and handles the logic of sending
 707      * snapshot chunks
 708      */
 709     protected class FollowerToSnapshot {
 710         private final ByteString snapshotBytes;
 711         private int offset = 0;
 712         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 713         private int replyReceivedForOffset;
 714         // if replyStatus is false, the previous chunk is attempted
 715         private boolean replyStatus = false;
 716         private int chunkIndex;
 717         private final int totalChunks;
 718         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 719         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 720
 721         public FollowerToSnapshot(ByteString snapshotBytes) {
 722             this.snapshotBytes = snapshotBytes;
 723             int size = snapshotBytes.size();
 724             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 725                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 726             if(LOG.isDebugEnabled()) {
 727                 LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}",
 728                         context.getId(), size, totalChunks);
 729             }
 730             replyReceivedForOffset = -1;
 731             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 732         }
 733
 734         public ByteString getSnapshotBytes() {
 735             return snapshotBytes;
 736         }
 737
 738         public int incrementOffset() {
 739             if(replyStatus) {
 740                 // if prev chunk failed, we would want to sent the same chunk again
 741                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 742             }
 743             return offset;
 744         }
 745
 746         public int incrementChunkIndex() {
 747             if (replyStatus) {
 748                 // if prev chunk failed, we would want to sent the same chunk again
 749                 chunkIndex =  chunkIndex + 1;
 750             }
 751             return chunkIndex;
 752         }
 753
 754         public int getChunkIndex() {
 755             return chunkIndex;
 756         }
 757
 758         public int getTotalChunks() {
 759             return totalChunks;
 760         }
 761
 762         public boolean canSendNextChunk() {
 763             // we only send a false if a chunk is sent but we have not received a reply yet
 764             return replyReceivedForOffset == offset;
 765         }
 766
 767         public boolean isLastChunk(int chunkIndex) {
 768             return totalChunks == chunkIndex;
 769         }
 770
 771         public void markSendStatus(boolean success) {
 772             if (success) {
 773                 // if the chunk sent was successful
 774                 replyReceivedForOffset = offset;
 775                 replyStatus = true;
 776                 lastChunkHashCode = nextChunkHashCode;
 777             } else {
 778                 // if the chunk sent was failure
 779                 replyReceivedForOffset = offset;
 780                 replyStatus = false;
 781             }
 782         }
 783
 784         public ByteString getNextChunk() {
 785             int snapshotLength = getSnapshotBytes().size();
 786             int start = incrementOffset();
 787             int size = context.getConfigParams().getSnapshotChunkSize();
 788             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 789                 size = snapshotLength;
 790             } else {
 791                 if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 792                     size = snapshotLength - start;
 793                 }
 794             }
 795
 796             if(LOG.isDebugEnabled()) {
 797                 LOG.debug("{}: Next chunk: length={}, offset={},size={}", context.getId(),
 798                     snapshotLength, start, size);
 799             }
 800             ByteString substring = getSnapshotBytes().substring(start, start + size);
 801             nextChunkHashCode = substring.hashCode();
 802             return substring;
 803         }
 804
 805         /**
 806          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 807          */
 808         public void reset(){
 809             offset = 0;
 810             replyStatus = false;
 811             replyReceivedForOffset = offset;
 812             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 813             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 814         }
 815
 816         public int getLastChunkHashCode() {
 817             return lastChunkHashCode;
 818         }
 819     }
 820
 821     // called from example-actor for printing the follower-states
 822     public String printFollowerStates() {
 823         final StringBuilder sb = new StringBuilder();
 824
 825         sb.append('[');
 826         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 827             sb.append('{');
 828             sb.append(followerLogInformation.getId());
 829             sb.append(" state:");
 830             sb.append(followerLogInformation.isFollowerActive());
 831             sb.append("},");
 832         }
 833         sb.append(']');
 834
 835         return sb.toString();
 836     }
 837
 838     @VisibleForTesting
 839     public FollowerLogInformation getFollower(String followerId) {
 840         return followerToLog.get(followerId);
 841     }
 842
 843     @VisibleForTesting
 844     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 845         mapFollowerToSnapshot.put(followerId, snapshot);
 846     }
 847
 848     @VisibleForTesting
 849     public int followerSnapshotSize() {
 850         return mapFollowerToSnapshot.size();
 851     }
 852
 853     @VisibleForTesting
 854     public int followerLogSize() {
 855         return followerToLog.size();
 856     }
 857 }