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