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 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.RaftActorContext;
  34 import org.opendaylight.controller.cluster.raft.RaftState;
  35 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  36 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
  37 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  38 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
  39 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  40 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  41 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  42 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  43 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  44 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  45 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  46 import scala.concurrent.duration.FiniteDuration;
  47
  48 /**
  49  * The behavior of a RaftActor when it is in the Leader state
  50  * <p/>
  51  * Leaders:
  52  * <ul>
  53  * <li> Upon election: send initial empty AppendEntries RPCs
  54  * (heartbeat) to each server; repeat during idle periods to
  55  * prevent election timeouts (§5.2)
  56  * <li> If command received from client: append entry to local log,
  57  * respond after entry applied to state machine (§5.3)
  58  * <li> If last log index ≥ nextIndex for a follower: send
  59  * AppendEntries RPC with log entries starting at nextIndex
  60  * <ul>
  61  * <li> If successful: update nextIndex and matchIndex for
  62  * follower (§5.3)
  63  * <li> If AppendEntries fails because of log inconsistency:
  64  * decrement nextIndex and retry (§5.3)
  65  * </ul>
  66  * <li> If there exists an N such that N > commitIndex, a majority
  67  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  68  * set commitIndex = N (§5.3, §5.4).
  69  */
  70 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
  71
  72     // The index of the first chunk that is sent when installing a snapshot
  73     public static final int FIRST_CHUNK_INDEX = 1;
  74
  75     // The index that the follower should respond with if it needs the install snapshot to be reset
  76     public static final int INVALID_CHUNK_INDEX = -1;
  77
  78     // This would be passed as the hash code of the last chunk when sending the first chunk
  79     public static final int INITIAL_LAST_CHUNK_HASH_CODE = -1;
  80
  81     private final Map<String, FollowerLogInformation> followerToLog;
  82     private final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
  83
  84     private Cancellable heartbeatSchedule = null;
  85
  86     private final Collection<ClientRequestTracker> trackerList = new LinkedList<>();
  87
  88     protected final int minReplicationCount;
  89
  90     protected final int minIsolatedLeaderPeerCount;
  91
  92     private Optional<ByteString> snapshot;
  93
  94     public AbstractLeader(RaftActorContext context) {
  95         super(context, RaftState.Leader);
  96
  97         final Builder<String, FollowerLogInformation> ftlBuilder = ImmutableMap.builder();
  98         for (String followerId : context.getPeerAddresses().keySet()) {
  99             FollowerLogInformation followerLogInformation =
 100                 new FollowerLogInformationImpl(followerId, -1, context);
 101
 102             ftlBuilder.put(followerId, followerLogInformation);
 103         }
 104         followerToLog = ftlBuilder.build();
 105
 106         leaderId = context.getId();
 107
 108         LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
 109
 110         minReplicationCount = getMajorityVoteCount(getFollowerIds().size());
 111
 112         // the isolated Leader peer count will be 1 less than the majority vote count.
 113         // this is because the vote count has the self vote counted in it
 114         // for e.g
 115         // 0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 116         // 2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 117         // 4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 118         minIsolatedLeaderPeerCount = minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 119
 120         snapshot = Optional.absent();
 121
 122         // Immediately schedule a heartbeat
 123         // Upon election: send initial empty AppendEntries RPCs
 124         // (heartbeat) to each server; repeat during idle periods to
 125         // prevent election timeouts (§5.2)
 126         sendAppendEntries(0, false);
 127
 128         // It is important to schedule this heartbeat here
 129         scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 130     }
 131
 132     /**
 133      * Return an immutable collection of follower identifiers.
 134      *
 135      * @return Collection of follower IDs
 136      */
 137     protected final Collection<String> getFollowerIds() {
 138         return followerToLog.keySet();
 139     }
 140
 141     @VisibleForTesting
 142     void setSnapshot(Optional<ByteString> snapshot) {
 143         this.snapshot = snapshot;
 144     }
 145
 146     @Override
 147     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 148         AppendEntries appendEntries) {
 149
 150         LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 151
 152         return this;
 153     }
 154
 155     @Override
 156     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 157         AppendEntriesReply appendEntriesReply) {
 158
 159         if(LOG.isTraceEnabled()) {
 160             LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
 161         }
 162
 163         // Update the FollowerLogInformation
 164         String followerId = appendEntriesReply.getFollowerId();
 165         FollowerLogInformation followerLogInformation =
 166             followerToLog.get(followerId);
 167
 168         if(followerLogInformation == null){
 169             LOG.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
 170             return this;
 171         }
 172
 173         if(followerLogInformation.timeSinceLastActivity() >
 174                 context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
 175             LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
 176                             "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
 177                     logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
 178                     context.getLastApplied(), context.getCommitIndex());
 179         }
 180
 181         followerLogInformation.markFollowerActive();
 182
 183         boolean updated = false;
 184         if (appendEntriesReply.isSuccess()) {
 185             updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
 186             updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
 187
 188             if(updated && LOG.isDebugEnabled()) {
 189                 LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}", logName(),
 190                         followerId, followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
 191             }
 192         } else {
 193             LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
 194
 195             // TODO: When we find that the follower is out of sync with the
 196             // Leader we simply decrement that followers next index by 1.
 197             // Would it be possible to do better than this? The RAFT spec
 198             // does not explicitly deal with it but may be something for us to
 199             // think about
 200
 201             followerLogInformation.decrNextIndex();
 202         }
 203
 204         // Now figure out if this reply warrants a change in the commitIndex
 205         // If there exists an N such that N > commitIndex, a majority
 206         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 207         // set commitIndex = N (§5.3, §5.4).
 208         for (long N = context.getCommitIndex() + 1; ; N++) {
 209             int replicatedCount = 1;
 210
 211             for (FollowerLogInformation info : followerToLog.values()) {
 212                 if (info.getMatchIndex() >= N) {
 213                     replicatedCount++;
 214                 }
 215             }
 216
 217             if (replicatedCount >= minReplicationCount) {
 218                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 219                 if (replicatedLogEntry != null &&
 220                     replicatedLogEntry.getTerm() == currentTerm()) {
 221                     context.setCommitIndex(N);
 222                 }
 223             } else {
 224                 break;
 225             }
 226         }
 227
 228         // Apply the change to the state machine
 229         if (context.getCommitIndex() > context.getLastApplied()) {
 230             if(LOG.isDebugEnabled()) {
 231                 LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
 232                         logName(), followerId, context.getCommitIndex(), context.getLastApplied());
 233             }
 234
 235             applyLogToStateMachine(context.getCommitIndex());
 236         }
 237
 238         if (!context.isSnapshotCaptureInitiated()) {
 239             purgeInMemoryLog();
 240         }
 241
 242         //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
 243         sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
 244         return this;
 245     }
 246
 247     private void purgeInMemoryLog() {
 248         //find the lowest index across followers which has been replicated to all.
 249         // lastApplied if there are no followers, so that we keep clearing the log for single-node
 250         // we would delete the in-mem log from that index on, in-order to minimize mem usage
 251         // we would also share this info thru AE with the followers so that they can delete their log entries as well.
 252         long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
 253         for (FollowerLogInformation info : followerToLog.values()) {
 254             minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
 255         }
 256
 257         super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
 258     }
 259
 260     @Override
 261     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 262         final Iterator<ClientRequestTracker> it = trackerList.iterator();
 263         while (it.hasNext()) {
 264             final ClientRequestTracker t = it.next();
 265             if (t.getIndex() == logIndex) {
 266                 it.remove();
 267                 return t;
 268             }
 269         }
 270
 271         return null;
 272     }
 273
 274     @Override
 275     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 276         for (ClientRequestTracker tracker : trackerList) {
 277             if (tracker.getIndex() == logIndex) {
 278                 return tracker;
 279             }
 280         }
 281         return null;
 282     }
 283
 284     @Override
 285     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 286         RequestVoteReply requestVoteReply) {
 287         return this;
 288     }
 289
 290     protected void beforeSendHeartbeat(){}
 291
 292     @Override
 293     public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 294         Preconditions.checkNotNull(sender, "sender should not be null");
 295
 296         Object message = fromSerializableMessage(originalMessage);
 297
 298         if (message instanceof RaftRPC) {
 299             RaftRPC rpc = (RaftRPC) message;
 300             // If RPC request or response contains term T > currentTerm:
 301             // set currentTerm = T, convert to follower (§5.1)
 302             // This applies to all RPC messages and responses
 303             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 304                 LOG.debug("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
 305                         logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 306
 307                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 308
 309                 return switchBehavior(new Follower(context));
 310             }
 311         }
 312
 313         if (message instanceof SendHeartBeat) {
 314             beforeSendHeartbeat();
 315             sendHeartBeat();
 316             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 317             return this;
 318
 319         } else if(message instanceof SendInstallSnapshot) {
 320             // received from RaftActor
 321             setSnapshot(Optional.of(((SendInstallSnapshot) message).getSnapshot()));
 322             sendInstallSnapshot();
 323
 324         } else if (message instanceof Replicate) {
 325             replicate((Replicate) message);
 326
 327         } else if (message instanceof InstallSnapshotReply){
 328             handleInstallSnapshotReply((InstallSnapshotReply) message);
 329
 330         }
 331
 332
 333         return super.handleMessage(sender, message);
 334     }
 335
 336     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 337         LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
 338
 339         String followerId = reply.getFollowerId();
 340         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 341
 342         if (followerToSnapshot == null) {
 343             LOG.error("{}: FollowerId {} in InstallSnapshotReply not known to Leader",
 344                     logName(), followerId);
 345             return;
 346         }
 347
 348         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 349         followerLogInformation.markFollowerActive();
 350
 351         if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 352             boolean wasLastChunk = false;
 353             if (reply.isSuccess()) {
 354                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 355                     //this was the last chunk reply
 356                     if(LOG.isDebugEnabled()) {
 357                         LOG.debug("{}: InstallSnapshotReply received, " +
 358                                 "last chunk received, Chunk: {}. Follower: {} Setting nextIndex: {}",
 359                                 logName(), reply.getChunkIndex(), followerId,
 360                             context.getReplicatedLog().getSnapshotIndex() + 1
 361                         );
 362                     }
 363
 364                     followerLogInformation.setMatchIndex(
 365                         context.getReplicatedLog().getSnapshotIndex());
 366                     followerLogInformation.setNextIndex(
 367                         context.getReplicatedLog().getSnapshotIndex() + 1);
 368                     mapFollowerToSnapshot.remove(followerId);
 369
 370                     LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}",
 371                                 logName(), followerId, followerLogInformation.getMatchIndex(),
 372                                 followerLogInformation.getNextIndex());
 373
 374                     if (mapFollowerToSnapshot.isEmpty()) {
 375                         // once there are no pending followers receiving snapshots
 376                         // we can remove snapshot from the memory
 377                         setSnapshot(Optional.<ByteString>absent());
 378                     }
 379                     wasLastChunk = true;
 380
 381                 } else {
 382                     followerToSnapshot.markSendStatus(true);
 383                 }
 384             } else {
 385                 LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
 386                         logName(), reply.getChunkIndex());
 387
 388                 followerToSnapshot.markSendStatus(false);
 389             }
 390
 391             if (wasLastChunk && !context.isSnapshotCaptureInitiated()) {
 392                 // Since the follower is now caught up try to purge the log.
 393                 purgeInMemoryLog();
 394             } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
 395                 ActorSelection followerActor = context.getPeerActorSelection(followerId);
 396                 if(followerActor != null) {
 397                     sendSnapshotChunk(followerActor, followerId);
 398                 }
 399             }
 400
 401         } else {
 402             LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
 403                     logName(), reply.getChunkIndex(), followerId,
 404                     followerToSnapshot.getChunkIndex());
 405
 406             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 407                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 408                 // so that Installing the snapshot can resume from the beginning
 409                 followerToSnapshot.reset();
 410             }
 411         }
 412     }
 413
 414     private void replicate(Replicate replicate) {
 415         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 416
 417         LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}", logName(),
 418                 replicate.getIdentifier(), logIndex);
 419
 420         // Create a tracker entry we will use this later to notify the
 421         // client actor
 422         trackerList.add(
 423             new ClientRequestTrackerImpl(replicate.getClientActor(),
 424                 replicate.getIdentifier(),
 425                 logIndex)
 426         );
 427
 428         if (followerToLog.isEmpty()) {
 429             context.setCommitIndex(logIndex);
 430             applyLogToStateMachine(logIndex);
 431         } else {
 432             sendAppendEntries(0, false);
 433         }
 434     }
 435
 436     private void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) {
 437         // Send an AppendEntries to all followers
 438         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 439             final String followerId = e.getKey();
 440             final FollowerLogInformation followerLogInformation = e.getValue();
 441             // This checks helps not to send a repeat message to the follower
 442             if(!followerLogInformation.isFollowerActive() ||
 443                     followerLogInformation.timeSinceLastActivity() >= timeSinceLastActivityInterval) {
 444                 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
 445             }
 446         }
 447     }
 448
 449     /**
 450      *
 451      * This method checks if any update needs to be sent to the given follower. This includes append log entries,
 452      * sending next snapshot chunk, and initiating a snapshot.
 453      * @return true if any update is sent, false otherwise
 454      */
 455
 456     private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
 457                                        boolean sendHeartbeat, boolean isHeartbeat) {
 458
 459         ActorSelection followerActor = context.getPeerActorSelection(followerId);
 460         if (followerActor != null) {
 461             long followerNextIndex = followerLogInformation.getNextIndex();
 462             boolean isFollowerActive = followerLogInformation.isFollowerActive();
 463
 464             if (mapFollowerToSnapshot.get(followerId) != null) {
 465                 // if install snapshot is in process , then sent next chunk if possible
 466                 if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 467                     sendSnapshotChunk(followerActor, followerId);
 468                 } else if(sendHeartbeat) {
 469                     // we send a heartbeat even if we have not received a reply for the last chunk
 470                     sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
 471                         Collections.<ReplicatedLogEntry>emptyList(), followerId);
 472                 }
 473             } else {
 474                 long leaderLastIndex = context.getReplicatedLog().lastIndex();
 475                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 476
 477                 if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
 478                     LOG.debug("{}: Checking sendAppendEntries for follower {}, followerNextIndex {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
 479                             logName(), followerId, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
 480                 }
 481
 482                 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
 483
 484                     LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
 485                             followerNextIndex, followerId);
 486
 487                     // FIXME : Sending one entry at a time
 488                     final List<ReplicatedLogEntry> entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
 489
 490                     sendAppendEntriesToFollower(followerActor, followerNextIndex, entries, followerId);
 491
 492                 } else if (isFollowerActive && followerNextIndex >= 0 &&
 493                     leaderLastIndex > followerNextIndex && !context.isSnapshotCaptureInitiated()) {
 494                     // if the followers next index is not present in the leaders log, and
 495                     // if the follower is just not starting and if leader's index is more than followers index
 496                     // then snapshot should be sent
 497
 498                     if (LOG.isDebugEnabled()) {
 499                         LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s," +
 500                                     "follower-nextIndex: %d, leader-snapshot-index: %d,  " +
 501                                     "leader-last-index: %d", logName(), followerId,
 502                                     followerNextIndex, leaderSnapShotIndex, leaderLastIndex));
 503                     }
 504
 505                     // Send heartbeat to follower whenever install snapshot is initiated.
 506                     sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
 507                             Collections.<ReplicatedLogEntry>emptyList(), followerId);
 508
 509                     initiateCaptureSnapshot(followerId, followerNextIndex);
 510
 511                 } else if(sendHeartbeat) {
 512                     //we send an AppendEntries, even if the follower is inactive
 513                     // in-order to update the followers timestamp, in case it becomes active again
 514                     sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
 515                         Collections.<ReplicatedLogEntry>emptyList(), followerId);
 516                 }
 517
 518             }
 519         }
 520     }
 521
 522     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 523         List<ReplicatedLogEntry> entries, String followerId) {
 524         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
 525             prevLogIndex(followerNextIndex),
 526             prevLogTerm(followerNextIndex), entries,
 527             context.getCommitIndex(), super.getReplicatedToAllIndex());
 528
 529         if(!entries.isEmpty() || LOG.isTraceEnabled()) {
 530             LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
 531                     appendEntries);
 532         }
 533
 534         followerActor.tell(appendEntries.toSerializable(), actor());
 535     }
 536
 537     /**
 538      * Install Snapshot works as follows
 539      * 1. Leader initiates the capture snapshot by sending a CaptureSnapshot message to actor
 540      * 2. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 541      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 542      * 3. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 543      * 4. On complete, Follower sends back a InstallSnapshotReply.
 544      * 5. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 545      * and replenishes the memory by deleting the snapshot in Replicated log.
 546      * 6. If another follower requires a snapshot and a snapshot has been collected (via CaptureSnapshotReply)
 547      * then send the existing snapshot in chunks to the follower.
 548      * @param followerId
 549      * @param followerNextIndex
 550      */
 551     private void initiateCaptureSnapshot(String followerId, long followerNextIndex) {
 552         if (!context.getReplicatedLog().isPresent(followerNextIndex) &&
 553                 context.getReplicatedLog().isInSnapshot(followerNextIndex)) {
 554
 555             if (snapshot.isPresent()) {
 556                 // if a snapshot is present in the memory, most likely another install is in progress
 557                 // no need to capture snapshot.
 558                 // This could happen if another follower needs an install when one is going on.
 559                 final ActorSelection followerActor = context.getPeerActorSelection(followerId);
 560                 sendSnapshotChunk(followerActor, followerId);
 561
 562             } else if (!context.isSnapshotCaptureInitiated()) {
 563
 564                 ReplicatedLogEntry lastAppliedEntry = context.getReplicatedLog().get(context.getLastApplied());
 565                 long lastAppliedIndex = -1;
 566                 long lastAppliedTerm = -1;
 567
 568                 if (lastAppliedEntry != null) {
 569                     lastAppliedIndex = lastAppliedEntry.getIndex();
 570                     lastAppliedTerm = lastAppliedEntry.getTerm();
 571                 } else if (context.getReplicatedLog().getSnapshotIndex() > -1) {
 572                     lastAppliedIndex = context.getReplicatedLog().getSnapshotIndex();
 573                     lastAppliedTerm = context.getReplicatedLog().getSnapshotTerm();
 574                 }
 575
 576                 boolean isInstallSnapshotInitiated = true;
 577                 long replicatedToAllIndex = super.getReplicatedToAllIndex();
 578                 ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
 579
 580                 CaptureSnapshot captureSnapshot = new CaptureSnapshot(
 581                         lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
 582                         (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
 583                         (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1),
 584                         isInstallSnapshotInitiated);
 585
 586                 if(LOG.isDebugEnabled()) {
 587                     LOG.debug("{}: Initiating install snapshot to follower {}: {}", logName(), followerId,
 588                             captureSnapshot);
 589                 }
 590
 591                 actor().tell(captureSnapshot, actor());
 592                 context.setSnapshotCaptureInitiated(true);
 593             }
 594         }
 595     }
 596
 597
 598     private void sendInstallSnapshot() {
 599         LOG.debug("{}: sendInstallSnapshot", logName());
 600         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 601             ActorSelection followerActor = context.getPeerActorSelection(e.getKey());
 602
 603             if (followerActor != null) {
 604                 long nextIndex = e.getValue().getNextIndex();
 605
 606                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 607                     context.getReplicatedLog().isInSnapshot(nextIndex)) {
 608                     sendSnapshotChunk(followerActor, e.getKey());
 609                 }
 610             }
 611         }
 612     }
 613
 614     /**
 615      *  Sends a snapshot chunk to a given follower
 616      *  InstallSnapshot should qualify as a heartbeat too.
 617      */
 618     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 619         try {
 620             if (snapshot.isPresent()) {
 621                 ByteString nextSnapshotChunk = getNextSnapshotChunk(followerId,snapshot.get());
 622
 623                 // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
 624                 // followerId to the followerToSnapshot map.
 625                 FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 626
 627                 followerActor.tell(
 628                     new InstallSnapshot(currentTerm(), context.getId(),
 629                         context.getReplicatedLog().getSnapshotIndex(),
 630                         context.getReplicatedLog().getSnapshotTerm(),
 631                         nextSnapshotChunk,
 632                         followerToSnapshot.incrementChunkIndex(),
 633                         followerToSnapshot.getTotalChunks(),
 634                         Optional.of(followerToSnapshot.getLastChunkHashCode())
 635                     ).toSerializable(),
 636                     actor()
 637                 );
 638
 639                 if(LOG.isDebugEnabled()) {
 640                     LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
 641                             logName(), followerActor.path(), followerToSnapshot.getChunkIndex(),
 642                             followerToSnapshot.getTotalChunks());
 643                 }
 644             }
 645         } catch (IOException e) {
 646             LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
 647         }
 648     }
 649
 650     /**
 651      * Acccepts snaphot as ByteString, enters into map for future chunks
 652      * creates and return a ByteString chunk
 653      */
 654     private ByteString getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 655         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 656         if (followerToSnapshot == null) {
 657             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 658             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 659         }
 660         ByteString nextChunk = followerToSnapshot.getNextChunk();
 661
 662         LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.size());
 663
 664         return nextChunk;
 665     }
 666
 667     private void sendHeartBeat() {
 668         if (!followerToLog.isEmpty()) {
 669             LOG.trace("{}: Sending heartbeat", logName());
 670             sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toMillis(), true);
 671         }
 672     }
 673
 674     private void stopHeartBeat() {
 675         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 676             heartbeatSchedule.cancel();
 677         }
 678     }
 679
 680     private void scheduleHeartBeat(FiniteDuration interval) {
 681         if (followerToLog.isEmpty()) {
 682             // Optimization - do not bother scheduling a heartbeat as there are
 683             // no followers
 684             return;
 685         }
 686
 687         stopHeartBeat();
 688
 689         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 690         // message is sent to itself.
 691         // Scheduling the heartbeat only once here because heartbeats do not
 692         // need to be sent if there are other messages being sent to the remote
 693         // actor.
 694         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 695             interval, context.getActor(), new SendHeartBeat(),
 696             context.getActorSystem().dispatcher(), context.getActor());
 697     }
 698
 699     @Override
 700     public void close() throws Exception {
 701         stopHeartBeat();
 702     }
 703
 704     @Override
 705     public String getLeaderId() {
 706         return context.getId();
 707     }
 708
 709     protected boolean isLeaderIsolated() {
 710         int minPresent = minIsolatedLeaderPeerCount;
 711         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 712             if (followerLogInformation.isFollowerActive()) {
 713                 --minPresent;
 714                 if (minPresent == 0) {
 715                     break;
 716                 }
 717             }
 718         }
 719         return (minPresent != 0);
 720     }
 721
 722     /**
 723      * Encapsulates the snapshot bytestring and handles the logic of sending
 724      * snapshot chunks
 725      */
 726     protected class FollowerToSnapshot {
 727         private final ByteString snapshotBytes;
 728         private int offset = 0;
 729         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 730         private int replyReceivedForOffset;
 731         // if replyStatus is false, the previous chunk is attempted
 732         private boolean replyStatus = false;
 733         private int chunkIndex;
 734         private final int totalChunks;
 735         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 736         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 737
 738         public FollowerToSnapshot(ByteString snapshotBytes) {
 739             this.snapshotBytes = snapshotBytes;
 740             int size = snapshotBytes.size();
 741             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 742                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 743             if(LOG.isDebugEnabled()) {
 744                 LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}",
 745                         logName(), size, totalChunks);
 746             }
 747             replyReceivedForOffset = -1;
 748             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 749         }
 750
 751         public ByteString getSnapshotBytes() {
 752             return snapshotBytes;
 753         }
 754
 755         public int incrementOffset() {
 756             if(replyStatus) {
 757                 // if prev chunk failed, we would want to sent the same chunk again
 758                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 759             }
 760             return offset;
 761         }
 762
 763         public int incrementChunkIndex() {
 764             if (replyStatus) {
 765                 // if prev chunk failed, we would want to sent the same chunk again
 766                 chunkIndex =  chunkIndex + 1;
 767             }
 768             return chunkIndex;
 769         }
 770
 771         public int getChunkIndex() {
 772             return chunkIndex;
 773         }
 774
 775         public int getTotalChunks() {
 776             return totalChunks;
 777         }
 778
 779         public boolean canSendNextChunk() {
 780             // we only send a false if a chunk is sent but we have not received a reply yet
 781             return replyReceivedForOffset == offset;
 782         }
 783
 784         public boolean isLastChunk(int chunkIndex) {
 785             return totalChunks == chunkIndex;
 786         }
 787
 788         public void markSendStatus(boolean success) {
 789             if (success) {
 790                 // if the chunk sent was successful
 791                 replyReceivedForOffset = offset;
 792                 replyStatus = true;
 793                 lastChunkHashCode = nextChunkHashCode;
 794             } else {
 795                 // if the chunk sent was failure
 796                 replyReceivedForOffset = offset;
 797                 replyStatus = false;
 798             }
 799         }
 800
 801         public ByteString getNextChunk() {
 802             int snapshotLength = getSnapshotBytes().size();
 803             int start = incrementOffset();
 804             int size = context.getConfigParams().getSnapshotChunkSize();
 805             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 806                 size = snapshotLength;
 807             } else {
 808                 if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 809                     size = snapshotLength - start;
 810                 }
 811             }
 812
 813
 814             LOG.debug("{}: Next chunk: length={}, offset={},size={}", logName(),
 815                     snapshotLength, start, size);
 816
 817             ByteString substring = getSnapshotBytes().substring(start, start + size);
 818             nextChunkHashCode = substring.hashCode();
 819             return substring;
 820         }
 821
 822         /**
 823          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 824          */
 825         public void reset(){
 826             offset = 0;
 827             replyStatus = false;
 828             replyReceivedForOffset = offset;
 829             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 830             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 831         }
 832
 833         public int getLastChunkHashCode() {
 834             return lastChunkHashCode;
 835         }
 836     }
 837
 838     // called from example-actor for printing the follower-states
 839     public String printFollowerStates() {
 840         final StringBuilder sb = new StringBuilder();
 841
 842         sb.append('[');
 843         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 844             sb.append('{');
 845             sb.append(followerLogInformation.getId());
 846             sb.append(" state:");
 847             sb.append(followerLogInformation.isFollowerActive());
 848             sb.append("},");
 849         }
 850         sb.append(']');
 851
 852         return sb.toString();
 853     }
 854
 855     @VisibleForTesting
 856     public FollowerLogInformation getFollower(String followerId) {
 857         return followerToLog.get(followerId);
 858     }
 859
 860     @VisibleForTesting
 861     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 862         mapFollowerToSnapshot.put(followerId, snapshot);
 863     }
 864
 865     @VisibleForTesting
 866     public int followerSnapshotSize() {
 867         return mapFollowerToSnapshot.size();
 868     }
 869
 870     @VisibleForTesting
 871     public int followerLogSize() {
 872         return followerToLog.size();
 873     }
 874 }