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 javax.annotation.Nullable;
  30 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
  31 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
  32 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
  33 import org.opendaylight.controller.cluster.raft.FollowerLogInformationImpl;
  34 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  35 import org.opendaylight.controller.cluster.raft.RaftState;
  36 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  37 import org.opendaylight.controller.cluster.raft.Snapshot;
  38 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  39 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
  40 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  41 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  42 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  43 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  44 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  45 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  46 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  47 import scala.concurrent.duration.FiniteDuration;
  48
  49 /**
  50  * The behavior of a RaftActor when it is in the Leader state
  51  * <p/>
  52  * Leaders:
  53  * <ul>
  54  * <li> Upon election: send initial empty AppendEntries RPCs
  55  * (heartbeat) to each server; repeat during idle periods to
  56  * prevent election timeouts (§5.2)
  57  * <li> If command received from client: append entry to local log,
  58  * respond after entry applied to state machine (§5.3)
  59  * <li> If last log index ≥ nextIndex for a follower: send
  60  * AppendEntries RPC with log entries starting at nextIndex
  61  * <ul>
  62  * <li> If successful: update nextIndex and matchIndex for
  63  * follower (§5.3)
  64  * <li> If AppendEntries fails because of log inconsistency:
  65  * decrement nextIndex and retry (§5.3)
  66  * </ul>
  67  * <li> If there exists an N such that N > commitIndex, a majority
  68  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  69  * set commitIndex = N (§5.3, §5.4).
  70  */
  71 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
  72
  73     // The index of the first chunk that is sent when installing a snapshot
  74     public static final int FIRST_CHUNK_INDEX = 1;
  75
  76     // The index that the follower should respond with if it needs the install snapshot to be reset
  77     public static final int INVALID_CHUNK_INDEX = -1;
  78
  79     // This would be passed as the hash code of the last chunk when sending the first chunk
  80     public static final int INITIAL_LAST_CHUNK_HASH_CODE = -1;
  81
  82     private final Map<String, FollowerLogInformation> followerToLog;
  83     private final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
  84
  85     private Cancellable heartbeatSchedule = null;
  86
  87     private final Collection<ClientRequestTracker> trackerList = new LinkedList<>();
  88
  89     protected final int minReplicationCount;
  90
  91     protected final int minIsolatedLeaderPeerCount;
  92
  93     private Optional<SnapshotHolder> snapshot;
  94
  95     public AbstractLeader(RaftActorContext context) {
  96         super(context, RaftState.Leader);
  97
  98         setLeaderPayloadVersion(context.getPayloadVersion());
  99
 100         final Builder<String, FollowerLogInformation> ftlBuilder = ImmutableMap.builder();
 101         for (String followerId : context.getPeerAddresses().keySet()) {
 102             FollowerLogInformation followerLogInformation =
 103                 new FollowerLogInformationImpl(followerId, -1, context);
 104
 105             ftlBuilder.put(followerId, followerLogInformation);
 106         }
 107         followerToLog = ftlBuilder.build();
 108
 109         leaderId = context.getId();
 110
 111         LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
 112
 113         minReplicationCount = getMajorityVoteCount(getFollowerIds().size());
 114
 115         // the isolated Leader peer count will be 1 less than the majority vote count.
 116         // this is because the vote count has the self vote counted in it
 117         // for e.g
 118         // 0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 119         // 2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 120         // 4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 121         minIsolatedLeaderPeerCount = minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 122
 123         snapshot = Optional.absent();
 124
 125         // Immediately schedule a heartbeat
 126         // Upon election: send initial empty AppendEntries RPCs
 127         // (heartbeat) to each server; repeat during idle periods to
 128         // prevent election timeouts (§5.2)
 129         sendAppendEntries(0, false);
 130
 131         // It is important to schedule this heartbeat here
 132         scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 133     }
 134
 135     /**
 136      * Return an immutable collection of follower identifiers.
 137      *
 138      * @return Collection of follower IDs
 139      */
 140     public final Collection<String> getFollowerIds() {
 141         return followerToLog.keySet();
 142     }
 143
 144     @VisibleForTesting
 145     void setSnapshot(@Nullable Snapshot snapshot) {
 146         if(snapshot != null) {
 147             this.snapshot = Optional.of(new SnapshotHolder(snapshot));
 148         } else {
 149             this.snapshot = Optional.absent();
 150         }
 151     }
 152
 153     @Override
 154     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 155         AppendEntries appendEntries) {
 156
 157         LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 158
 159         return this;
 160     }
 161
 162     @Override
 163     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 164         AppendEntriesReply appendEntriesReply) {
 165
 166         if(LOG.isTraceEnabled()) {
 167             LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
 168         }
 169
 170         // Update the FollowerLogInformation
 171         String followerId = appendEntriesReply.getFollowerId();
 172         FollowerLogInformation followerLogInformation =
 173             followerToLog.get(followerId);
 174
 175         if(followerLogInformation == null){
 176             LOG.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
 177             return this;
 178         }
 179
 180         if(followerLogInformation.timeSinceLastActivity() >
 181                 context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
 182             LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
 183                             "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
 184                     logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
 185                     context.getLastApplied(), context.getCommitIndex());
 186         }
 187
 188         followerLogInformation.markFollowerActive();
 189         followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
 190
 191         boolean updated = false;
 192         if (appendEntriesReply.isSuccess()) {
 193             updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 194         } else {
 195             LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
 196
 197             long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
 198             ReplicatedLogEntry followersLastLogEntry = context.getReplicatedLog().get(followerLastLogIndex);
 199             if(followerLastLogIndex < 0 || (followersLastLogEntry != null &&
 200                     followersLastLogEntry.getTerm() == appendEntriesReply.getLogLastTerm())) {
 201                 // The follower's log is empty or the last entry is present in the leader's journal
 202                 // and the terms match so the follower is just behind the leader's journal from
 203                 // the last snapshot, if any. We'll catch up the follower quickly by starting at the
 204                 // follower's last log index.
 205
 206                 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 207             } else {
 208                 // TODO: When we find that the follower is out of sync with the
 209                 // Leader we simply decrement that followers next index by 1.
 210                 // Would it be possible to do better than this? The RAFT spec
 211                 // does not explicitly deal with it but may be something for us to
 212                 // think about.
 213
 214                 followerLogInformation.decrNextIndex();
 215             }
 216         }
 217
 218         // Now figure out if this reply warrants a change in the commitIndex
 219         // If there exists an N such that N > commitIndex, a majority
 220         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 221         // set commitIndex = N (§5.3, §5.4).
 222         for (long N = context.getCommitIndex() + 1; ; N++) {
 223             int replicatedCount = 1;
 224
 225             for (FollowerLogInformation info : followerToLog.values()) {
 226                 if (info.getMatchIndex() >= N) {
 227                     replicatedCount++;
 228                 }
 229             }
 230
 231             if (replicatedCount >= minReplicationCount) {
 232                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 233                 if (replicatedLogEntry != null &&
 234                     replicatedLogEntry.getTerm() == currentTerm()) {
 235                     context.setCommitIndex(N);
 236                 }
 237             } else {
 238                 break;
 239             }
 240         }
 241
 242         // Apply the change to the state machine
 243         if (context.getCommitIndex() > context.getLastApplied()) {
 244             if(LOG.isDebugEnabled()) {
 245                 LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
 246                         logName(), followerId, context.getCommitIndex(), context.getLastApplied());
 247             }
 248
 249             applyLogToStateMachine(context.getCommitIndex());
 250         }
 251
 252         if (!context.getSnapshotManager().isCapturing()) {
 253             purgeInMemoryLog();
 254         }
 255
 256         //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
 257         sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
 258         return this;
 259     }
 260
 261     private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
 262             AppendEntriesReply appendEntriesReply) {
 263         boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
 264         updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
 265
 266         if(updated && LOG.isDebugEnabled()) {
 267             LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
 268                     logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
 269                     followerLogInformation.getNextIndex());
 270         }
 271         return updated;
 272     }
 273
 274     private void purgeInMemoryLog() {
 275         //find the lowest index across followers which has been replicated to all.
 276         // lastApplied if there are no followers, so that we keep clearing the log for single-node
 277         // we would delete the in-mem log from that index on, in-order to minimize mem usage
 278         // we would also share this info thru AE with the followers so that they can delete their log entries as well.
 279         long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
 280         for (FollowerLogInformation info : followerToLog.values()) {
 281             minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
 282         }
 283
 284         super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
 285     }
 286
 287     @Override
 288     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 289         final Iterator<ClientRequestTracker> it = trackerList.iterator();
 290         while (it.hasNext()) {
 291             final ClientRequestTracker t = it.next();
 292             if (t.getIndex() == logIndex) {
 293                 it.remove();
 294                 return t;
 295             }
 296         }
 297
 298         return null;
 299     }
 300
 301     @Override
 302     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 303         for (ClientRequestTracker tracker : trackerList) {
 304             if (tracker.getIndex() == logIndex) {
 305                 return tracker;
 306             }
 307         }
 308         return null;
 309     }
 310
 311     @Override
 312     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 313         RequestVoteReply requestVoteReply) {
 314         return this;
 315     }
 316
 317     protected void beforeSendHeartbeat(){}
 318
 319     @Override
 320     public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 321         Preconditions.checkNotNull(sender, "sender should not be null");
 322
 323         Object message = fromSerializableMessage(originalMessage);
 324
 325         if (message instanceof RaftRPC) {
 326             RaftRPC rpc = (RaftRPC) message;
 327             // If RPC request or response contains term T > currentTerm:
 328             // set currentTerm = T, convert to follower (§5.1)
 329             // This applies to all RPC messages and responses
 330             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 331                 LOG.debug("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
 332                         logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 333
 334                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 335
 336                 return switchBehavior(new Follower(context));
 337             }
 338         }
 339
 340         if (message instanceof SendHeartBeat) {
 341             beforeSendHeartbeat();
 342             sendHeartBeat();
 343             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 344             return this;
 345
 346         } else if(message instanceof SendInstallSnapshot) {
 347             // received from RaftActor
 348             setSnapshot(((SendInstallSnapshot) message).getSnapshot());
 349             sendInstallSnapshot();
 350
 351         } else if (message instanceof Replicate) {
 352             replicate((Replicate) message);
 353
 354         } else if (message instanceof InstallSnapshotReply){
 355             handleInstallSnapshotReply((InstallSnapshotReply) message);
 356
 357         }
 358
 359
 360         return super.handleMessage(sender, message);
 361     }
 362
 363     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 364         LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
 365
 366         String followerId = reply.getFollowerId();
 367         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 368
 369         if (followerToSnapshot == null) {
 370             LOG.error("{}: FollowerId {} in InstallSnapshotReply not known to Leader",
 371                     logName(), followerId);
 372             return;
 373         }
 374
 375         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 376         followerLogInformation.markFollowerActive();
 377
 378         if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 379             boolean wasLastChunk = false;
 380             if (reply.isSuccess()) {
 381                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 382                     //this was the last chunk reply
 383                     if(LOG.isDebugEnabled()) {
 384                         LOG.debug("{}: InstallSnapshotReply received, " +
 385                                 "last chunk received, Chunk: {}. Follower: {} Setting nextIndex: {}",
 386                                 logName(), reply.getChunkIndex(), followerId,
 387                             context.getReplicatedLog().getSnapshotIndex() + 1
 388                         );
 389                     }
 390
 391                     long followerMatchIndex = snapshot.get().getLastIncludedIndex();
 392                     followerLogInformation.setMatchIndex(followerMatchIndex);
 393                     followerLogInformation.setNextIndex(followerMatchIndex + 1);
 394                     mapFollowerToSnapshot.remove(followerId);
 395
 396                     LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}",
 397                                 logName(), followerId, followerLogInformation.getMatchIndex(),
 398                                 followerLogInformation.getNextIndex());
 399
 400                     if (mapFollowerToSnapshot.isEmpty()) {
 401                         // once there are no pending followers receiving snapshots
 402                         // we can remove snapshot from the memory
 403                         setSnapshot(null);
 404                     }
 405                     wasLastChunk = true;
 406
 407                 } else {
 408                     followerToSnapshot.markSendStatus(true);
 409                 }
 410             } else {
 411                 LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
 412                         logName(), reply.getChunkIndex());
 413
 414                 followerToSnapshot.markSendStatus(false);
 415             }
 416
 417             if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
 418                 // Since the follower is now caught up try to purge the log.
 419                 purgeInMemoryLog();
 420             } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
 421                 ActorSelection followerActor = context.getPeerActorSelection(followerId);
 422                 if(followerActor != null) {
 423                     sendSnapshotChunk(followerActor, followerId);
 424                 }
 425             }
 426
 427         } else {
 428             LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
 429                     logName(), reply.getChunkIndex(), followerId,
 430                     followerToSnapshot.getChunkIndex());
 431
 432             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 433                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 434                 // so that Installing the snapshot can resume from the beginning
 435                 followerToSnapshot.reset();
 436             }
 437         }
 438     }
 439
 440     private void replicate(Replicate replicate) {
 441         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 442
 443         LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}", logName(),
 444                 replicate.getIdentifier(), logIndex);
 445
 446         // Create a tracker entry we will use this later to notify the
 447         // client actor
 448         trackerList.add(
 449             new ClientRequestTrackerImpl(replicate.getClientActor(),
 450                 replicate.getIdentifier(),
 451                 logIndex)
 452         );
 453
 454         boolean applyModificationToState = followerToLog.isEmpty()
 455                 || context.getRaftPolicy().applyModificationToStateBeforeConsensus();
 456
 457         if(applyModificationToState){
 458             context.setCommitIndex(logIndex);
 459             applyLogToStateMachine(logIndex);
 460         }
 461
 462         if (!followerToLog.isEmpty()) {
 463             sendAppendEntries(0, false);
 464         }
 465     }
 466
 467     private void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) {
 468         // Send an AppendEntries to all followers
 469         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 470             final String followerId = e.getKey();
 471             final FollowerLogInformation followerLogInformation = e.getValue();
 472             // This checks helps not to send a repeat message to the follower
 473             if(!followerLogInformation.isFollowerActive() ||
 474                     followerLogInformation.timeSinceLastActivity() >= timeSinceLastActivityInterval) {
 475                 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
 476             }
 477         }
 478     }
 479
 480     /**
 481      *
 482      * This method checks if any update needs to be sent to the given follower. This includes append log entries,
 483      * sending next snapshot chunk, and initiating a snapshot.
 484      * @return true if any update is sent, false otherwise
 485      */
 486
 487     private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
 488                                        boolean sendHeartbeat, boolean isHeartbeat) {
 489
 490         ActorSelection followerActor = context.getPeerActorSelection(followerId);
 491         if (followerActor != null) {
 492             long followerNextIndex = followerLogInformation.getNextIndex();
 493             boolean isFollowerActive = followerLogInformation.isFollowerActive();
 494             boolean sendAppendEntries = false;
 495             List<ReplicatedLogEntry> entries = Collections.emptyList();
 496
 497             if (mapFollowerToSnapshot.get(followerId) != null) {
 498                 // if install snapshot is in process , then sent next chunk if possible
 499                 if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 500                     sendSnapshotChunk(followerActor, followerId);
 501                 } else if(sendHeartbeat) {
 502                     // we send a heartbeat even if we have not received a reply for the last chunk
 503                     sendAppendEntries = true;
 504                 }
 505             } else {
 506                 long leaderLastIndex = context.getReplicatedLog().lastIndex();
 507                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 508
 509                 if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
 510                     LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
 511                             logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
 512                 }
 513
 514                 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
 515
 516                     LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
 517                             followerNextIndex, followerId);
 518
 519                     if(followerLogInformation.okToReplicate()) {
 520                         // Try to send all the entries in the journal but not exceeding the max data size
 521                         // for a single AppendEntries message.
 522                         int maxEntries = (int) context.getReplicatedLog().size();
 523                         entries = context.getReplicatedLog().getFrom(followerNextIndex, maxEntries,
 524                                 context.getConfigParams().getSnapshotChunkSize());
 525                         sendAppendEntries = true;
 526                     }
 527                 } else if (isFollowerActive && followerNextIndex >= 0 &&
 528                     leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
 529                     // if the followers next index is not present in the leaders log, and
 530                     // if the follower is just not starting and if leader's index is more than followers index
 531                     // then snapshot should be sent
 532
 533                     if (LOG.isDebugEnabled()) {
 534                         LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s," +
 535                                     "follower-nextIndex: %d, leader-snapshot-index: %d,  " +
 536                                     "leader-last-index: %d", logName(), followerId,
 537                                     followerNextIndex, leaderSnapShotIndex, leaderLastIndex));
 538                     }
 539
 540                     // Send heartbeat to follower whenever install snapshot is initiated.
 541                     sendAppendEntries = true;
 542                     initiateCaptureSnapshot(followerId, followerNextIndex);
 543
 544                 } else if(sendHeartbeat) {
 545                     // we send an AppendEntries, even if the follower is inactive
 546                     // in-order to update the followers timestamp, in case it becomes active again
 547                     sendAppendEntries = true;
 548                 }
 549
 550             }
 551
 552             if(sendAppendEntries) {
 553                 sendAppendEntriesToFollower(followerActor, followerNextIndex,
 554                         entries, followerId);
 555             }
 556         }
 557     }
 558
 559     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 560         List<ReplicatedLogEntry> entries, String followerId) {
 561         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
 562             prevLogIndex(followerNextIndex),
 563             prevLogTerm(followerNextIndex), entries,
 564             context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
 565
 566         if(!entries.isEmpty() || LOG.isTraceEnabled()) {
 567             LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
 568                     appendEntries);
 569         }
 570
 571         followerActor.tell(appendEntries.toSerializable(), actor());
 572     }
 573
 574     /**
 575      * Install Snapshot works as follows
 576      * 1. Leader initiates the capture snapshot by sending a CaptureSnapshot message to actor
 577      * 2. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 578      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 579      * 3. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 580      * 4. On complete, Follower sends back a InstallSnapshotReply.
 581      * 5. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 582      * and replenishes the memory by deleting the snapshot in Replicated log.
 583      * 6. If another follower requires a snapshot and a snapshot has been collected (via CaptureSnapshotReply)
 584      * then send the existing snapshot in chunks to the follower.
 585      * @param followerId
 586      * @param followerNextIndex
 587      */
 588     private void initiateCaptureSnapshot(String followerId, long followerNextIndex) {
 589         if (!context.getReplicatedLog().isPresent(followerNextIndex) &&
 590                 context.getReplicatedLog().isInSnapshot(followerNextIndex)) {
 591
 592             if (snapshot.isPresent()) {
 593                 // if a snapshot is present in the memory, most likely another install is in progress
 594                 // no need to capture snapshot.
 595                 // This could happen if another follower needs an install when one is going on.
 596                 final ActorSelection followerActor = context.getPeerActorSelection(followerId);
 597                 sendSnapshotChunk(followerActor, followerId);
 598
 599
 600             } else {
 601                 context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
 602                         this.getReplicatedToAllIndex(), followerId);
 603             }
 604         }
 605     }
 606
 607
 608     private void sendInstallSnapshot() {
 609         LOG.debug("{}: sendInstallSnapshot", logName());
 610         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 611             ActorSelection followerActor = context.getPeerActorSelection(e.getKey());
 612
 613             if (followerActor != null) {
 614                 long nextIndex = e.getValue().getNextIndex();
 615
 616                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 617                     context.getReplicatedLog().isInSnapshot(nextIndex)) {
 618                     sendSnapshotChunk(followerActor, e.getKey());
 619                 }
 620             }
 621         }
 622     }
 623
 624     /**
 625      *  Sends a snapshot chunk to a given follower
 626      *  InstallSnapshot should qualify as a heartbeat too.
 627      */
 628     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 629         try {
 630             if (snapshot.isPresent()) {
 631                 ByteString nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes());
 632
 633                 // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
 634                 // followerId to the followerToSnapshot map.
 635                 FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 636
 637                 followerActor.tell(
 638                     new InstallSnapshot(currentTerm(), context.getId(),
 639                         snapshot.get().getLastIncludedIndex(),
 640                         snapshot.get().getLastIncludedTerm(),
 641                         nextSnapshotChunk,
 642                         followerToSnapshot.incrementChunkIndex(),
 643                         followerToSnapshot.getTotalChunks(),
 644                         Optional.of(followerToSnapshot.getLastChunkHashCode())
 645                     ).toSerializable(),
 646                     actor()
 647                 );
 648
 649                 if(LOG.isDebugEnabled()) {
 650                     LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
 651                             logName(), followerActor.path(), followerToSnapshot.getChunkIndex(),
 652                             followerToSnapshot.getTotalChunks());
 653                 }
 654             }
 655         } catch (IOException e) {
 656             LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
 657         }
 658     }
 659
 660     /**
 661      * Acccepts snaphot as ByteString, enters into map for future chunks
 662      * creates and return a ByteString chunk
 663      */
 664     private ByteString getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 665         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 666         if (followerToSnapshot == null) {
 667             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 668             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 669         }
 670         ByteString nextChunk = followerToSnapshot.getNextChunk();
 671
 672         LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.size());
 673
 674         return nextChunk;
 675     }
 676
 677     private void sendHeartBeat() {
 678         if (!followerToLog.isEmpty()) {
 679             LOG.trace("{}: Sending heartbeat", logName());
 680             sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toMillis(), true);
 681         }
 682     }
 683
 684     private void stopHeartBeat() {
 685         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 686             heartbeatSchedule.cancel();
 687         }
 688     }
 689
 690     private void scheduleHeartBeat(FiniteDuration interval) {
 691         if (followerToLog.isEmpty()) {
 692             // Optimization - do not bother scheduling a heartbeat as there are
 693             // no followers
 694             return;
 695         }
 696
 697         stopHeartBeat();
 698
 699         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 700         // message is sent to itself.
 701         // Scheduling the heartbeat only once here because heartbeats do not
 702         // need to be sent if there are other messages being sent to the remote
 703         // actor.
 704         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 705             interval, context.getActor(), new SendHeartBeat(),
 706             context.getActorSystem().dispatcher(), context.getActor());
 707     }
 708
 709     @Override
 710     public void close() throws Exception {
 711         stopHeartBeat();
 712     }
 713
 714     @Override
 715     public String getLeaderId() {
 716         return context.getId();
 717     }
 718
 719     protected boolean isLeaderIsolated() {
 720         int minPresent = minIsolatedLeaderPeerCount;
 721         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 722             if (followerLogInformation.isFollowerActive()) {
 723                 --minPresent;
 724                 if (minPresent == 0) {
 725                     break;
 726                 }
 727             }
 728         }
 729         return (minPresent != 0);
 730     }
 731
 732     /**
 733      * Encapsulates the snapshot bytestring and handles the logic of sending
 734      * snapshot chunks
 735      */
 736     protected class FollowerToSnapshot {
 737         private final ByteString snapshotBytes;
 738         private int offset = 0;
 739         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 740         private int replyReceivedForOffset;
 741         // if replyStatus is false, the previous chunk is attempted
 742         private boolean replyStatus = false;
 743         private int chunkIndex;
 744         private final int totalChunks;
 745         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 746         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 747
 748         public FollowerToSnapshot(ByteString snapshotBytes) {
 749             this.snapshotBytes = snapshotBytes;
 750             int size = snapshotBytes.size();
 751             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 752                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 753             if(LOG.isDebugEnabled()) {
 754                 LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}",
 755                         logName(), size, totalChunks);
 756             }
 757             replyReceivedForOffset = -1;
 758             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 759         }
 760
 761         public ByteString getSnapshotBytes() {
 762             return snapshotBytes;
 763         }
 764
 765         public int incrementOffset() {
 766             if(replyStatus) {
 767                 // if prev chunk failed, we would want to sent the same chunk again
 768                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 769             }
 770             return offset;
 771         }
 772
 773         public int incrementChunkIndex() {
 774             if (replyStatus) {
 775                 // if prev chunk failed, we would want to sent the same chunk again
 776                 chunkIndex =  chunkIndex + 1;
 777             }
 778             return chunkIndex;
 779         }
 780
 781         public int getChunkIndex() {
 782             return chunkIndex;
 783         }
 784
 785         public int getTotalChunks() {
 786             return totalChunks;
 787         }
 788
 789         public boolean canSendNextChunk() {
 790             // we only send a false if a chunk is sent but we have not received a reply yet
 791             return replyReceivedForOffset == offset;
 792         }
 793
 794         public boolean isLastChunk(int chunkIndex) {
 795             return totalChunks == chunkIndex;
 796         }
 797
 798         public void markSendStatus(boolean success) {
 799             if (success) {
 800                 // if the chunk sent was successful
 801                 replyReceivedForOffset = offset;
 802                 replyStatus = true;
 803                 lastChunkHashCode = nextChunkHashCode;
 804             } else {
 805                 // if the chunk sent was failure
 806                 replyReceivedForOffset = offset;
 807                 replyStatus = false;
 808             }
 809         }
 810
 811         public ByteString getNextChunk() {
 812             int snapshotLength = getSnapshotBytes().size();
 813             int start = incrementOffset();
 814             int size = context.getConfigParams().getSnapshotChunkSize();
 815             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 816                 size = snapshotLength;
 817             } else {
 818                 if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 819                     size = snapshotLength - start;
 820                 }
 821             }
 822
 823
 824             LOG.debug("{}: Next chunk: length={}, offset={},size={}", logName(),
 825                     snapshotLength, start, size);
 826
 827             ByteString substring = getSnapshotBytes().substring(start, start + size);
 828             nextChunkHashCode = substring.hashCode();
 829             return substring;
 830         }
 831
 832         /**
 833          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 834          */
 835         public void reset(){
 836             offset = 0;
 837             replyStatus = false;
 838             replyReceivedForOffset = offset;
 839             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 840             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 841         }
 842
 843         public int getLastChunkHashCode() {
 844             return lastChunkHashCode;
 845         }
 846     }
 847
 848     // called from example-actor for printing the follower-states
 849     public String printFollowerStates() {
 850         final StringBuilder sb = new StringBuilder();
 851
 852         sb.append('[');
 853         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 854             sb.append('{');
 855             sb.append(followerLogInformation.getId());
 856             sb.append(" state:");
 857             sb.append(followerLogInformation.isFollowerActive());
 858             sb.append("},");
 859         }
 860         sb.append(']');
 861
 862         return sb.toString();
 863     }
 864
 865     @VisibleForTesting
 866     public FollowerLogInformation getFollower(String followerId) {
 867         return followerToLog.get(followerId);
 868     }
 869
 870     @VisibleForTesting
 871     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 872         mapFollowerToSnapshot.put(followerId, snapshot);
 873     }
 874
 875     @VisibleForTesting
 876     public int followerSnapshotSize() {
 877         return mapFollowerToSnapshot.size();
 878     }
 879
 880     @VisibleForTesting
 881     public int followerLogSize() {
 882         return followerToLog.size();
 883     }
 884
 885     private static class SnapshotHolder {
 886         private final long lastIncludedTerm;
 887         private final long lastIncludedIndex;
 888         private final ByteString snapshotBytes;
 889
 890         SnapshotHolder(Snapshot snapshot) {
 891             this.lastIncludedTerm = snapshot.getLastAppliedTerm();
 892             this.lastIncludedIndex = snapshot.getLastAppliedIndex();
 893             this.snapshotBytes = ByteString.copyFrom(snapshot.getState());
 894         }
 895
 896         long getLastIncludedTerm() {
 897             return lastIncludedTerm;
 898         }
 899
 900         long getLastIncludedIndex() {
 901             return lastIncludedIndex;
 902         }
 903
 904         ByteString getSnapshotBytes() {
 905             return snapshotBytes;
 906         }
 907     }
 908 }