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