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