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.protobuf.ByteString;
  18 import java.io.IOException;
  19 import java.util.Arrays;
  20 import java.util.Collection;
  21 import java.util.Collections;
  22 import java.util.HashMap;
  23 import java.util.Iterator;
  24 import java.util.LinkedList;
  25 import java.util.List;
  26 import java.util.Map;
  27 import java.util.Map.Entry;
  28 import javax.annotation.Nullable;
  29 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
  30 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
  31 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
  32 import org.opendaylight.controller.cluster.raft.FollowerLogInformationImpl;
  33 import org.opendaylight.controller.cluster.raft.PeerInfo;
  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.Snapshot;
  38 import org.opendaylight.controller.cluster.raft.VotingState;
  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 org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply;
  49 import scala.concurrent.duration.FiniteDuration;
  50
  51 /**
  52  * The behavior of a RaftActor when it is in the Leader state
  53  * <p/>
  54  * Leaders:
  55  * <ul>
  56  * <li> Upon election: send initial empty AppendEntries RPCs
  57  * (heartbeat) to each server; repeat during idle periods to
  58  * prevent election timeouts (§5.2)
  59  * <li> If command received from client: append entry to local log,
  60  * respond after entry applied to state machine (§5.3)
  61  * <li> If last log index ≥ nextIndex for a follower: send
  62  * AppendEntries RPC with log entries starting at nextIndex
  63  * <ul>
  64  * <li> If successful: update nextIndex and matchIndex for
  65  * follower (§5.3)
  66  * <li> If AppendEntries fails because of log inconsistency:
  67  * decrement nextIndex and retry (§5.3)
  68  * </ul>
  69  * <li> If there exists an N such that N > commitIndex, a majority
  70  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  71  * set commitIndex = N (§5.3, §5.4).
  72  */
  73 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
  74
  75     // The index of the first chunk that is sent when installing a snapshot
  76     public static final int FIRST_CHUNK_INDEX = 1;
  77
  78     // The index that the follower should respond with if it needs the install snapshot to be reset
  79     public static final int INVALID_CHUNK_INDEX = -1;
  80
  81     // This would be passed as the hash code of the last chunk when sending the first chunk
  82     public static final int INITIAL_LAST_CHUNK_HASH_CODE = -1;
  83
  84     private final Map<String, FollowerLogInformation> followerToLog = new HashMap<>();
  85     private final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
  86
  87     private Cancellable heartbeatSchedule = null;
  88
  89     private final Collection<ClientRequestTracker> trackerList = new LinkedList<>();
  90
  91     private int minReplicationCount;
  92
  93     private Optional<SnapshotHolder> snapshot;
  94
  95     protected AbstractLeader(RaftActorContext context, RaftState state) {
  96         super(context, state);
  97
  98         for(PeerInfo peerInfo: context.getPeers()) {
  99             FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(peerInfo, -1, context);
 100             followerToLog.put(peerInfo.getId(), followerLogInformation);
 101         }
 102
 103         LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
 104
 105         updateMinReplicaCount();
 106
 107         snapshot = Optional.absent();
 108
 109         // Immediately schedule a heartbeat
 110         // Upon election: send initial empty AppendEntries RPCs
 111         // (heartbeat) to each server; repeat during idle periods to
 112         // prevent election timeouts (§5.2)
 113         sendAppendEntries(0, false);
 114
 115         // It is important to schedule this heartbeat here
 116         scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 117     }
 118
 119     /**
 120      * Return an immutable collection of follower identifiers.
 121      *
 122      * @return Collection of follower IDs
 123      */
 124     public final Collection<String> getFollowerIds() {
 125         return followerToLog.keySet();
 126     }
 127
 128     public void addFollower(String followerId) {
 129         FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(
 130                 context.getPeerInfo(followerId), -1, context);
 131         followerToLog.put(followerId, followerLogInformation);
 132
 133         if(heartbeatSchedule == null) {
 134             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 135         }
 136     }
 137
 138     public void removeFollower(String followerId) {
 139         followerToLog.remove(followerId);
 140         mapFollowerToSnapshot.remove(followerId);
 141     }
 142
 143     public void updateMinReplicaCount() {
 144         int numVoting = 0;
 145         for(PeerInfo peer: context.getPeers()) {
 146             if(peer.isVoting()) {
 147                 numVoting++;
 148             }
 149         }
 150
 151         minReplicationCount = getMajorityVoteCount(numVoting);
 152     }
 153
 154     protected int getMinIsolatedLeaderPeerCount(){
 155       //the isolated Leader peer count will be 1 less than the majority vote count.
 156         //this is because the vote count has the self vote counted in it
 157         //for e.g
 158         //0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 159         //2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 160         //4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 161
 162         return minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 163     }
 164
 165     @VisibleForTesting
 166     void setSnapshot(@Nullable Snapshot snapshot) {
 167         if(snapshot != null) {
 168             this.snapshot = Optional.of(new SnapshotHolder(snapshot));
 169         } else {
 170             this.snapshot = Optional.absent();
 171         }
 172     }
 173
 174     @Override
 175     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 176         AppendEntries appendEntries) {
 177
 178         LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 179
 180         return this;
 181     }
 182
 183     @Override
 184     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 185         AppendEntriesReply appendEntriesReply) {
 186
 187         if(LOG.isTraceEnabled()) {
 188             LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
 189         }
 190
 191         // Update the FollowerLogInformation
 192         String followerId = appendEntriesReply.getFollowerId();
 193         FollowerLogInformation followerLogInformation =
 194             followerToLog.get(followerId);
 195
 196         if(followerLogInformation == null){
 197             LOG.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
 198             return this;
 199         }
 200
 201         if(followerLogInformation.timeSinceLastActivity() >
 202                 context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
 203             LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
 204                             "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
 205                     logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
 206                     context.getLastApplied(), context.getCommitIndex());
 207         }
 208
 209         followerLogInformation.markFollowerActive();
 210         followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
 211         followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
 212
 213         boolean updated = false;
 214         if (appendEntriesReply.isSuccess()) {
 215             updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 216         } else {
 217             LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
 218
 219             long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
 220             ReplicatedLogEntry followersLastLogEntry = context.getReplicatedLog().get(followerLastLogIndex);
 221             if(appendEntriesReply.isForceInstallSnapshot()) {
 222                 // Reset the followers match and next index. This is to signal that this follower has nothing
 223                 // in common with this Leader and so would require a snapshot to be installed
 224                 followerLogInformation.setMatchIndex(-1);
 225                 followerLogInformation.setNextIndex(-1);
 226
 227                 // Force initiate a snapshot capture
 228                 initiateCaptureSnapshot(followerId);
 229             } else if(followerLastLogIndex < 0 || (followersLastLogEntry != null &&
 230                     followersLastLogEntry.getTerm() == appendEntriesReply.getLogLastTerm())) {
 231                 // The follower's log is empty or the last entry is present in the leader's journal
 232                 // and the terms match so the follower is just behind the leader's journal from
 233                 // the last snapshot, if any. We'll catch up the follower quickly by starting at the
 234                 // follower's last log index.
 235
 236                 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 237             } else {
 238                 // TODO: When we find that the follower is out of sync with the
 239                 // Leader we simply decrement that followers next index by 1.
 240                 // Would it be possible to do better than this? The RAFT spec
 241                 // does not explicitly deal with it but may be something for us to
 242                 // think about.
 243
 244                 followerLogInformation.decrNextIndex();
 245             }
 246         }
 247
 248         // Now figure out if this reply warrants a change in the commitIndex
 249         // If there exists an N such that N > commitIndex, a majority
 250         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 251         // set commitIndex = N (§5.3, §5.4).
 252         for (long N = context.getCommitIndex() + 1; ; N++) {
 253             int replicatedCount = 1;
 254
 255             for (FollowerLogInformation info : followerToLog.values()) {
 256                 final PeerInfo peerInfo = context.getPeerInfo(info.getId());
 257                 if(info.getMatchIndex() >= N && (peerInfo != null && peerInfo.isVoting())) {
 258                     replicatedCount++;
 259                 }
 260             }
 261
 262             if (replicatedCount >= minReplicationCount) {
 263                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 264                 if (replicatedLogEntry == null) {
 265                     break;
 266                 }
 267
 268                 // Don't update the commit index if the log entry is from a previous term, as per §5.4.1:
 269                 // "Raft never commits log entries from previous terms by counting replicas".
 270                 // However we keep looping so we can make progress when new entries in the current term
 271                 // reach consensus, as per §5.4.1: "once an entry from the current term is committed by
 272                 // counting replicas, then all prior entries are committed indirectly".
 273                 if (replicatedLogEntry.getTerm() == currentTerm()) {
 274                     context.setCommitIndex(N);
 275                 }
 276             } else {
 277                 break;
 278             }
 279         }
 280
 281         // Apply the change to the state machine
 282         if (context.getCommitIndex() > context.getLastApplied()) {
 283             if(LOG.isDebugEnabled()) {
 284                 LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
 285                         logName(), followerId, context.getCommitIndex(), context.getLastApplied());
 286             }
 287
 288             applyLogToStateMachine(context.getCommitIndex());
 289         }
 290
 291         if (!context.getSnapshotManager().isCapturing()) {
 292             purgeInMemoryLog();
 293         }
 294
 295         //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
 296         sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
 297         return this;
 298     }
 299
 300     private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
 301             AppendEntriesReply appendEntriesReply) {
 302         boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
 303         updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
 304
 305         if(updated && LOG.isDebugEnabled()) {
 306             LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
 307                     logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
 308                     followerLogInformation.getNextIndex());
 309         }
 310         return updated;
 311     }
 312
 313     private void purgeInMemoryLog() {
 314         //find the lowest index across followers which has been replicated to all.
 315         // lastApplied if there are no followers, so that we keep clearing the log for single-node
 316         // we would delete the in-mem log from that index on, in-order to minimize mem usage
 317         // we would also share this info thru AE with the followers so that they can delete their log entries as well.
 318         long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
 319         for (FollowerLogInformation info : followerToLog.values()) {
 320             minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
 321         }
 322
 323         super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
 324     }
 325
 326     @Override
 327     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 328         final Iterator<ClientRequestTracker> it = trackerList.iterator();
 329         while (it.hasNext()) {
 330             final ClientRequestTracker t = it.next();
 331             if (t.getIndex() == logIndex) {
 332                 it.remove();
 333                 return t;
 334             }
 335         }
 336
 337         return null;
 338     }
 339
 340     @Override
 341     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 342         for (ClientRequestTracker tracker : trackerList) {
 343             if (tracker.getIndex() == logIndex) {
 344                 return tracker;
 345             }
 346         }
 347         return null;
 348     }
 349
 350     @Override
 351     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 352         RequestVoteReply requestVoteReply) {
 353         return this;
 354     }
 355
 356     protected void beforeSendHeartbeat(){}
 357
 358     @Override
 359     public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 360         Preconditions.checkNotNull(sender, "sender should not be null");
 361
 362         Object message = fromSerializableMessage(originalMessage);
 363
 364         if (message instanceof RaftRPC) {
 365             RaftRPC rpc = (RaftRPC) message;
 366             // If RPC request or response contains term T > currentTerm:
 367             // set currentTerm = T, convert to follower (§5.1)
 368             // This applies to all RPC messages and responses
 369             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 370                 LOG.debug("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
 371                         logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 372
 373                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 374
 375                 return internalSwitchBehavior(RaftState.Follower);
 376             }
 377         }
 378
 379         if (message instanceof SendHeartBeat) {
 380             beforeSendHeartbeat();
 381             sendHeartBeat();
 382             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 383             return this;
 384
 385         } else if(message instanceof SendInstallSnapshot) {
 386             // received from RaftActor
 387             setSnapshot(((SendInstallSnapshot) message).getSnapshot());
 388             sendInstallSnapshot();
 389
 390         } else if (message instanceof Replicate) {
 391             replicate((Replicate) message);
 392
 393         } else if (message instanceof InstallSnapshotReply){
 394             handleInstallSnapshotReply((InstallSnapshotReply) message);
 395
 396         }
 397
 398
 399         return super.handleMessage(sender, message);
 400     }
 401
 402     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 403         LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
 404
 405         String followerId = reply.getFollowerId();
 406         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 407
 408         if (followerToSnapshot == null) {
 409             LOG.error("{}: FollowerToSnapshot not found for follower {} in InstallSnapshotReply",
 410                     logName(), followerId);
 411             return;
 412         }
 413
 414         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 415         if(followerLogInformation == null) {
 416             // This can happen during AddServer if it times out.
 417             LOG.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
 418                     logName(), followerId);
 419             mapFollowerToSnapshot.remove(followerId);
 420             return;
 421         }
 422
 423         followerLogInformation.markFollowerActive();
 424
 425         if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 426             boolean wasLastChunk = false;
 427             if (reply.isSuccess()) {
 428                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 429                     //this was the last chunk reply
 430                     if(LOG.isDebugEnabled()) {
 431                         LOG.debug("{}: InstallSnapshotReply received, " +
 432                                 "last chunk received, Chunk: {}. Follower: {} Setting nextIndex: {}",
 433                                 logName(), reply.getChunkIndex(), followerId,
 434                             context.getReplicatedLog().getSnapshotIndex() + 1
 435                         );
 436                     }
 437
 438                     long followerMatchIndex = snapshot.get().getLastIncludedIndex();
 439                     followerLogInformation.setMatchIndex(followerMatchIndex);
 440                     followerLogInformation.setNextIndex(followerMatchIndex + 1);
 441                     mapFollowerToSnapshot.remove(followerId);
 442
 443                     LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}",
 444                         logName(), followerId, followerLogInformation.getMatchIndex(),
 445                         followerLogInformation.getNextIndex());
 446
 447                     if (mapFollowerToSnapshot.isEmpty()) {
 448                         // once there are no pending followers receiving snapshots
 449                         // we can remove snapshot from the memory
 450                         setSnapshot(null);
 451                     }
 452                     wasLastChunk = true;
 453                     if(context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED){
 454                         UnInitializedFollowerSnapshotReply unInitFollowerSnapshotSuccess =
 455                                              new UnInitializedFollowerSnapshotReply(followerId);
 456                         context.getActor().tell(unInitFollowerSnapshotSuccess, context.getActor());
 457                         LOG.debug("Sent message UnInitializedFollowerSnapshotReply to self");
 458                     }
 459                 } else {
 460                     followerToSnapshot.markSendStatus(true);
 461                 }
 462             } else {
 463                 LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
 464                         logName(), reply.getChunkIndex());
 465
 466                 followerToSnapshot.markSendStatus(false);
 467             }
 468
 469             if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
 470                 // Since the follower is now caught up try to purge the log.
 471                 purgeInMemoryLog();
 472             } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
 473                 ActorSelection followerActor = context.getPeerActorSelection(followerId);
 474                 if(followerActor != null) {
 475                     sendSnapshotChunk(followerActor, followerId);
 476                 }
 477             }
 478
 479         } else {
 480             LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
 481                     logName(), reply.getChunkIndex(), followerId,
 482                     followerToSnapshot.getChunkIndex());
 483
 484             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 485                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 486                 // so that Installing the snapshot can resume from the beginning
 487                 followerToSnapshot.reset();
 488             }
 489         }
 490     }
 491
 492     private void replicate(Replicate replicate) {
 493         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 494
 495         LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}", logName(),
 496                 replicate.getIdentifier(), logIndex);
 497
 498         // Create a tracker entry we will use this later to notify the
 499         // client actor
 500         trackerList.add(
 501             new ClientRequestTrackerImpl(replicate.getClientActor(),
 502                 replicate.getIdentifier(),
 503                 logIndex)
 504         );
 505
 506         boolean applyModificationToState = followerToLog.isEmpty()
 507                 || context.getRaftPolicy().applyModificationToStateBeforeConsensus();
 508
 509         if(applyModificationToState){
 510             context.setCommitIndex(logIndex);
 511             applyLogToStateMachine(logIndex);
 512         }
 513
 514         if (!followerToLog.isEmpty()) {
 515             sendAppendEntries(0, false);
 516         }
 517     }
 518
 519     protected void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) {
 520         // Send an AppendEntries to all followers
 521         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 522             final String followerId = e.getKey();
 523             final FollowerLogInformation followerLogInformation = e.getValue();
 524             // This checks helps not to send a repeat message to the follower
 525             if(!followerLogInformation.isFollowerActive() ||
 526                     followerLogInformation.timeSinceLastActivity() >= timeSinceLastActivityInterval) {
 527                 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
 528             }
 529         }
 530     }
 531
 532     /**
 533      *
 534      * This method checks if any update needs to be sent to the given follower. This includes append log entries,
 535      * sending next snapshot chunk, and initiating a snapshot.
 536      * @return true if any update is sent, false otherwise
 537      */
 538
 539     private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
 540                                        boolean sendHeartbeat, boolean isHeartbeat) {
 541
 542         ActorSelection followerActor = context.getPeerActorSelection(followerId);
 543         if (followerActor != null) {
 544             long followerNextIndex = followerLogInformation.getNextIndex();
 545             boolean isFollowerActive = followerLogInformation.isFollowerActive();
 546             boolean sendAppendEntries = false;
 547             List<ReplicatedLogEntry> entries = Collections.emptyList();
 548
 549             if (mapFollowerToSnapshot.get(followerId) != null) {
 550                 // if install snapshot is in process , then sent next chunk if possible
 551                 if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 552                     sendSnapshotChunk(followerActor, followerId);
 553                 } else if(sendHeartbeat) {
 554                     // we send a heartbeat even if we have not received a reply for the last chunk
 555                     sendAppendEntries = true;
 556                 }
 557             } else {
 558                 long leaderLastIndex = context.getReplicatedLog().lastIndex();
 559                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 560
 561                 if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
 562                     LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
 563                             logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
 564                 }
 565
 566                 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
 567
 568                     LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
 569                             followerNextIndex, followerId);
 570
 571                     if(followerLogInformation.okToReplicate()) {
 572                         // Try to send all the entries in the journal but not exceeding the max data size
 573                         // for a single AppendEntries message.
 574                         int maxEntries = (int) context.getReplicatedLog().size();
 575                         entries = context.getReplicatedLog().getFrom(followerNextIndex, maxEntries,
 576                                 context.getConfigParams().getSnapshotChunkSize());
 577                         sendAppendEntries = true;
 578                     }
 579                 } else if (isFollowerActive && followerNextIndex >= 0 &&
 580                     leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
 581                     // if the followers next index is not present in the leaders log, and
 582                     // if the follower is just not starting and if leader's index is more than followers index
 583                     // then snapshot should be sent
 584
 585                     if (LOG.isDebugEnabled()) {
 586                         LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s," +
 587                                     "follower-nextIndex: %d, leader-snapshot-index: %d,  " +
 588                                     "leader-last-index: %d", logName(), followerId,
 589                                     followerNextIndex, leaderSnapShotIndex, leaderLastIndex));
 590                     }
 591
 592                     // Send heartbeat to follower whenever install snapshot is initiated.
 593                     sendAppendEntries = true;
 594                     if (canInstallSnapshot(followerNextIndex)) {
 595                         initiateCaptureSnapshot(followerId);
 596                     }
 597
 598                 } else if(sendHeartbeat) {
 599                     // we send an AppendEntries, even if the follower is inactive
 600                     // in-order to update the followers timestamp, in case it becomes active again
 601                     sendAppendEntries = true;
 602                 }
 603
 604             }
 605
 606             if(sendAppendEntries) {
 607                 sendAppendEntriesToFollower(followerActor, followerNextIndex,
 608                         entries, followerId);
 609             }
 610         }
 611     }
 612
 613     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 614         List<ReplicatedLogEntry> entries, String followerId) {
 615         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
 616             prevLogIndex(followerNextIndex),
 617             prevLogTerm(followerNextIndex), entries,
 618             context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
 619
 620         if(!entries.isEmpty() || LOG.isTraceEnabled()) {
 621             LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
 622                     appendEntries);
 623         }
 624
 625         followerActor.tell(appendEntries, actor());
 626     }
 627
 628     /**
 629      * Install Snapshot works as follows
 630      * 1. Leader initiates the capture snapshot by sending a CaptureSnapshot message to actor
 631      * 2. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 632      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 633      * 3. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 634      * 4. On complete, Follower sends back a InstallSnapshotReply.
 635      * 5. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 636      * and replenishes the memory by deleting the snapshot in Replicated log.
 637      * 6. If another follower requires a snapshot and a snapshot has been collected (via CaptureSnapshotReply)
 638      * then send the existing snapshot in chunks to the follower.
 639      * @param followerId
 640      */
 641     public boolean initiateCaptureSnapshot(String followerId) {
 642         if (snapshot.isPresent()) {
 643             // if a snapshot is present in the memory, most likely another install is in progress
 644             // no need to capture snapshot.
 645             // This could happen if another follower needs an install when one is going on.
 646             final ActorSelection followerActor = context.getPeerActorSelection(followerId);
 647             sendSnapshotChunk(followerActor, followerId);
 648             return true;
 649         } else {
 650             return context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
 651                     this.getReplicatedToAllIndex(), followerId);
 652         }
 653     }
 654
 655     private boolean canInstallSnapshot(long nextIndex){
 656         // If the follower's nextIndex is -1 then we might as well send it a snapshot
 657         // Otherwise send it a snapshot only if the nextIndex is not present in the log but is present
 658         // in the snapshot
 659         return (nextIndex == -1 ||
 660                 (!context.getReplicatedLog().isPresent(nextIndex)
 661                         && context.getReplicatedLog().isInSnapshot(nextIndex)));
 662
 663     }
 664
 665
 666     private void sendInstallSnapshot() {
 667         LOG.debug("{}: sendInstallSnapshot", logName());
 668         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 669             String followerId = e.getKey();
 670             ActorSelection followerActor = context.getPeerActorSelection(followerId);
 671             FollowerLogInformation followerLogInfo = e.getValue();
 672
 673             if (followerActor != null) {
 674                 long nextIndex = followerLogInfo.getNextIndex();
 675                 if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED ||
 676                         canInstallSnapshot(nextIndex)) {
 677                     sendSnapshotChunk(followerActor, followerId);
 678                 }
 679             }
 680         }
 681     }
 682
 683     /**
 684      *  Sends a snapshot chunk to a given follower
 685      *  InstallSnapshot should qualify as a heartbeat too.
 686      */
 687     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 688         try {
 689             if (snapshot.isPresent()) {
 690                 byte[] nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes());
 691
 692                 // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
 693                 // followerId to the followerToSnapshot map.
 694                 FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 695
 696                 followerActor.tell(
 697                     new InstallSnapshot(currentTerm(), context.getId(),
 698                         snapshot.get().getLastIncludedIndex(),
 699                         snapshot.get().getLastIncludedTerm(),
 700                         nextSnapshotChunk,
 701                         followerToSnapshot.incrementChunkIndex(),
 702                         followerToSnapshot.getTotalChunks(),
 703                         Optional.of(followerToSnapshot.getLastChunkHashCode())
 704                     ).toSerializable(followerToLog.get(followerId).getRaftVersion()),
 705                     actor()
 706                 );
 707
 708                 if(LOG.isDebugEnabled()) {
 709                     LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
 710                             logName(), followerActor.path(), followerToSnapshot.getChunkIndex(),
 711                             followerToSnapshot.getTotalChunks());
 712                 }
 713             }
 714         } catch (IOException e) {
 715             LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
 716         }
 717     }
 718
 719     /**
 720      * Acccepts snaphot as ByteString, enters into map for future chunks
 721      * creates and return a ByteString chunk
 722      */
 723     private byte[] getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 724         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 725         if (followerToSnapshot == null) {
 726             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 727             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 728         }
 729         byte[] nextChunk = followerToSnapshot.getNextChunk();
 730
 731         LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.length);
 732
 733         return nextChunk;
 734     }
 735
 736     private void sendHeartBeat() {
 737         if (!followerToLog.isEmpty()) {
 738             LOG.trace("{}: Sending heartbeat", logName());
 739             sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toMillis(), true);
 740         }
 741     }
 742
 743     private void stopHeartBeat() {
 744         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 745             heartbeatSchedule.cancel();
 746         }
 747     }
 748
 749     private void scheduleHeartBeat(FiniteDuration interval) {
 750         if (followerToLog.isEmpty()) {
 751             // Optimization - do not bother scheduling a heartbeat as there are
 752             // no followers
 753             return;
 754         }
 755
 756         stopHeartBeat();
 757
 758         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 759         // message is sent to itself.
 760         // Scheduling the heartbeat only once here because heartbeats do not
 761         // need to be sent if there are other messages being sent to the remote
 762         // actor.
 763         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 764             interval, context.getActor(), SendHeartBeat.INSTANCE,
 765             context.getActorSystem().dispatcher(), context.getActor());
 766     }
 767
 768     @Override
 769     public void close() {
 770         stopHeartBeat();
 771     }
 772
 773     @Override
 774     public final String getLeaderId() {
 775         return context.getId();
 776     }
 777
 778     @Override
 779     public final short getLeaderPayloadVersion() {
 780         return context.getPayloadVersion();
 781     }
 782
 783     protected boolean isLeaderIsolated() {
 784         int minPresent = getMinIsolatedLeaderPeerCount();
 785         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 786             if (followerLogInformation.isFollowerActive()) {
 787                 --minPresent;
 788                 if (minPresent == 0) {
 789                     break;
 790                 }
 791             }
 792         }
 793         return (minPresent != 0);
 794     }
 795
 796     /**
 797      * Encapsulates the snapshot bytestring and handles the logic of sending
 798      * snapshot chunks
 799      */
 800     protected class FollowerToSnapshot {
 801         private final ByteString snapshotBytes;
 802         private int offset = 0;
 803         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 804         private int replyReceivedForOffset;
 805         // if replyStatus is false, the previous chunk is attempted
 806         private boolean replyStatus = false;
 807         private int chunkIndex;
 808         private final int totalChunks;
 809         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 810         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 811
 812         public FollowerToSnapshot(ByteString snapshotBytes) {
 813             this.snapshotBytes = snapshotBytes;
 814             int size = snapshotBytes.size();
 815             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 816                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 817             if(LOG.isDebugEnabled()) {
 818                 LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}",
 819                         logName(), size, totalChunks);
 820             }
 821             replyReceivedForOffset = -1;
 822             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 823         }
 824
 825         public ByteString getSnapshotBytes() {
 826             return snapshotBytes;
 827         }
 828
 829         public int incrementOffset() {
 830             if(replyStatus) {
 831                 // if prev chunk failed, we would want to sent the same chunk again
 832                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 833             }
 834             return offset;
 835         }
 836
 837         public int incrementChunkIndex() {
 838             if (replyStatus) {
 839                 // if prev chunk failed, we would want to sent the same chunk again
 840                 chunkIndex =  chunkIndex + 1;
 841             }
 842             return chunkIndex;
 843         }
 844
 845         public int getChunkIndex() {
 846             return chunkIndex;
 847         }
 848
 849         public int getTotalChunks() {
 850             return totalChunks;
 851         }
 852
 853         public boolean canSendNextChunk() {
 854             // we only send a false if a chunk is sent but we have not received a reply yet
 855             return replyReceivedForOffset == offset;
 856         }
 857
 858         public boolean isLastChunk(int chunkIndex) {
 859             return totalChunks == chunkIndex;
 860         }
 861
 862         public void markSendStatus(boolean success) {
 863             if (success) {
 864                 // if the chunk sent was successful
 865                 replyReceivedForOffset = offset;
 866                 replyStatus = true;
 867                 lastChunkHashCode = nextChunkHashCode;
 868             } else {
 869                 // if the chunk sent was failure
 870                 replyReceivedForOffset = offset;
 871                 replyStatus = false;
 872             }
 873         }
 874
 875         public byte[] getNextChunk() {
 876             int snapshotLength = getSnapshotBytes().size();
 877             int start = incrementOffset();
 878             int size = context.getConfigParams().getSnapshotChunkSize();
 879             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 880                 size = snapshotLength;
 881             } else if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 882                 size = snapshotLength - start;
 883             }
 884
 885             byte[] nextChunk = new byte[size];
 886             getSnapshotBytes().copyTo(nextChunk, start, 0, size);
 887             nextChunkHashCode = Arrays.hashCode(nextChunk);
 888
 889             LOG.debug("{}: Next chunk: total length={}, offset={}, size={}, hashCode={}", logName(),
 890                     snapshotLength, start, size, nextChunkHashCode);
 891             return nextChunk;
 892         }
 893
 894         /**
 895          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 896          */
 897         public void reset(){
 898             offset = 0;
 899             replyStatus = false;
 900             replyReceivedForOffset = offset;
 901             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 902             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 903         }
 904
 905         public int getLastChunkHashCode() {
 906             return lastChunkHashCode;
 907         }
 908     }
 909
 910     // called from example-actor for printing the follower-states
 911     public String printFollowerStates() {
 912         final StringBuilder sb = new StringBuilder();
 913
 914         sb.append('[');
 915         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 916             sb.append('{');
 917             sb.append(followerLogInformation.getId());
 918             sb.append(" state:");
 919             sb.append(followerLogInformation.isFollowerActive());
 920             sb.append("},");
 921         }
 922         sb.append(']');
 923
 924         return sb.toString();
 925     }
 926
 927     @VisibleForTesting
 928     public FollowerLogInformation getFollower(String followerId) {
 929         return followerToLog.get(followerId);
 930     }
 931
 932     @VisibleForTesting
 933     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 934         mapFollowerToSnapshot.put(followerId, snapshot);
 935     }
 936
 937     @VisibleForTesting
 938     public int followerSnapshotSize() {
 939         return mapFollowerToSnapshot.size();
 940     }
 941
 942     @VisibleForTesting
 943     public int followerLogSize() {
 944         return followerToLog.size();
 945     }
 946
 947     private static class SnapshotHolder {
 948         private final long lastIncludedTerm;
 949         private final long lastIncludedIndex;
 950         private final ByteString snapshotBytes;
 951
 952         SnapshotHolder(Snapshot snapshot) {
 953             this.lastIncludedTerm = snapshot.getLastAppliedTerm();
 954             this.lastIncludedIndex = snapshot.getLastAppliedIndex();
 955             this.snapshotBytes = ByteString.copyFrom(snapshot.getState());
 956         }
 957
 958         long getLastIncludedTerm() {
 959             return lastIncludedTerm;
 960         }
 961
 962         long getLastIncludedIndex() {
 963             return lastIncludedIndex;
 964         }
 965
 966         ByteString getSnapshotBytes() {
 967             return snapshotBytes;
 968         }
 969     }
 970 }