opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java

   1 /*
   2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
   3  *
   4  * This program and the accompanying materials are made available under the
   5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   6  * and is available at http://www.eclipse.org/legal/epl-v10.html
   7  */
   8
   9 package org.opendaylight.controller.cluster.raft.behaviors;
  10
  11 import akka.actor.ActorRef;
  12 import akka.actor.ActorSelection;
  13 import akka.actor.Cancellable;
  14 import com.google.common.annotations.VisibleForTesting;
  15 import com.google.common.base.Optional;
  16 import com.google.common.base.Preconditions;
  17 import com.google.protobuf.ByteString;
  18 import java.io.IOException;
  19 import java.util.Arrays;
  20 import java.util.Collection;
  21 import java.util.Collections;
  22 import java.util.HashMap;
  23 import java.util.Iterator;
  24 import java.util.LinkedList;
  25 import java.util.List;
  26 import java.util.Map;
  27 import java.util.Map.Entry;
  28 import java.util.Queue;
  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.PeerInfo;
  35 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  36 import org.opendaylight.controller.cluster.raft.RaftState;
  37 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  38 import org.opendaylight.controller.cluster.raft.ServerConfigurationPayload;
  39 import org.opendaylight.controller.cluster.raft.Snapshot;
  40 import org.opendaylight.controller.cluster.raft.VotingState;
  41 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  42 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
  43 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  44 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  45 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  46 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  47 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  48 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  49 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  50 import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply;
  51 import scala.concurrent.duration.FiniteDuration;
  52
  53 /**
  54  * The behavior of a RaftActor when it is in the Leader state
  55  * <p/>
  56  * Leaders:
  57  * <ul>
  58  * <li> Upon election: send initial empty AppendEntries RPCs
  59  * (heartbeat) to each server; repeat during idle periods to
  60  * prevent election timeouts (§5.2)
  61  * <li> If command received from client: append entry to local log,
  62  * respond after entry applied to state machine (§5.3)
  63  * <li> If last log index ≥ nextIndex for a follower: send
  64  * AppendEntries RPC with log entries starting at nextIndex
  65  * <ul>
  66  * <li> If successful: update nextIndex and matchIndex for
  67  * follower (§5.3)
  68  * <li> If AppendEntries fails because of log inconsistency:
  69  * decrement nextIndex and retry (§5.3)
  70  * </ul>
  71  * <li> If there exists an N such that N > commitIndex, a majority
  72  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  73  * set commitIndex = N (§5.3, §5.4).
  74  */
  75 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
  76
  77     // The index of the first chunk that is sent when installing a snapshot
  78     public static final int FIRST_CHUNK_INDEX = 1;
  79
  80     // The index that the follower should respond with if it needs the install snapshot to be reset
  81     public static final int INVALID_CHUNK_INDEX = -1;
  82
  83     // This would be passed as the hash code of the last chunk when sending the first chunk
  84     public static final int INITIAL_LAST_CHUNK_HASH_CODE = -1;
  85
  86     private final Map<String, FollowerLogInformation> followerToLog = new HashMap<>();
  87     private final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
  88
  89     /**
  90      * Lookup table for request contexts based on journal index. We could use a {@link Map} here, but we really
  91      * expect the entries to be modified in sequence, hence we open-code the lookup.
  92      *
  93      * TODO: Evaluate the use of ArrayDeque(), as that has lower memory overhead. Non-head removals are more costly,
  94      *       but we already expect those to be far from frequent.
  95      */
  96     private final Queue<ClientRequestTracker> trackers = new LinkedList<>();
  97
  98     private Cancellable heartbeatSchedule = null;
  99     private Optional<SnapshotHolder> snapshot;
 100     private int minReplicationCount;
 101
 102     protected AbstractLeader(RaftActorContext context, RaftState state) {
 103         super(context, state);
 104
 105         for(PeerInfo peerInfo: context.getPeers()) {
 106             FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(peerInfo, -1, context);
 107             followerToLog.put(peerInfo.getId(), followerLogInformation);
 108         }
 109
 110         LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
 111
 112         updateMinReplicaCount();
 113
 114         snapshot = Optional.absent();
 115
 116         // Immediately schedule a heartbeat
 117         // Upon election: send initial empty AppendEntries RPCs
 118         // (heartbeat) to each server; repeat during idle periods to
 119         // prevent election timeouts (§5.2)
 120         sendAppendEntries(0, false);
 121
 122         // It is important to schedule this heartbeat here
 123         scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 124     }
 125
 126     /**
 127      * Return an immutable collection of follower identifiers.
 128      *
 129      * @return Collection of follower IDs
 130      */
 131     public final Collection<String> getFollowerIds() {
 132         return followerToLog.keySet();
 133     }
 134
 135     public void addFollower(String followerId) {
 136         FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(
 137                 context.getPeerInfo(followerId), -1, context);
 138         followerToLog.put(followerId, followerLogInformation);
 139
 140         if(heartbeatSchedule == null) {
 141             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 142         }
 143     }
 144
 145     public void removeFollower(String followerId) {
 146         followerToLog.remove(followerId);
 147         mapFollowerToSnapshot.remove(followerId);
 148     }
 149
 150     public void updateMinReplicaCount() {
 151         int numVoting = 0;
 152         for(PeerInfo peer: context.getPeers()) {
 153             if(peer.isVoting()) {
 154                 numVoting++;
 155             }
 156         }
 157
 158         minReplicationCount = getMajorityVoteCount(numVoting);
 159     }
 160
 161     protected int getMinIsolatedLeaderPeerCount(){
 162       //the isolated Leader peer count will be 1 less than the majority vote count.
 163         //this is because the vote count has the self vote counted in it
 164         //for e.g
 165         //0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 166         //2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 167         //4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 168
 169         return minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 170     }
 171
 172     @VisibleForTesting
 173     void setSnapshot(@Nullable Snapshot snapshot) {
 174         if(snapshot != null) {
 175             this.snapshot = Optional.of(new SnapshotHolder(snapshot));
 176         } else {
 177             this.snapshot = Optional.absent();
 178         }
 179     }
 180
 181     @Override
 182     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 183         AppendEntries appendEntries) {
 184
 185         LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 186
 187         return this;
 188     }
 189
 190     @Override
 191     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 192         AppendEntriesReply appendEntriesReply) {
 193
 194         if(LOG.isTraceEnabled()) {
 195             LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
 196         }
 197
 198         // Update the FollowerLogInformation
 199         String followerId = appendEntriesReply.getFollowerId();
 200         FollowerLogInformation followerLogInformation =
 201             followerToLog.get(followerId);
 202
 203         if(followerLogInformation == null){
 204             LOG.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
 205             return this;
 206         }
 207
 208         if(followerLogInformation.timeSinceLastActivity() >
 209                 context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
 210             LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
 211                             "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
 212                     logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
 213                     context.getLastApplied(), context.getCommitIndex());
 214         }
 215
 216         followerLogInformation.markFollowerActive();
 217         followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
 218         followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
 219
 220         boolean updated = false;
 221         if(appendEntriesReply.getLogLastIndex() > context.getReplicatedLog().lastIndex()) {
 222             // The follower's log is actually ahead of the leader's log. Normally this doesn't happen
 223             // in raft as a node cannot become leader if it's log is behind another's. However, the
 224             // non-voting semantics deviate a bit from raft. Only voting members participate in
 225             // elections and can become leader so it's possible for a non-voting follower to be ahead
 226             // of the leader. This can happen if persistence is disabled and all voting members are
 227             // restarted. In this case, the voting leader will start out with an empty log however
 228             // the non-voting followers still retain the previous data in memory. On the first
 229             // AppendEntries, the non-voting follower returns a successful reply b/c the prevLogIndex
 230             // sent by the leader is -1 and thus the integrity checks pass. However the follower's returned
 231             // lastLogIndex may be higher in which case we want to reset the follower by installing a
 232             // snapshot. It's also possible that the follower's last log index is behind the leader's.
 233             // However in this case the log terms won't match and the logs will conflict - this is handled
 234             // elsewhere.
 235             LOG.debug("{}: handleAppendEntriesReply: follower {} lastIndex {} is ahead of our lastIndex {} - forcing install snaphot",
 236                     logName(), followerLogInformation.getId(), appendEntriesReply.getLogLastIndex(),
 237                     context.getReplicatedLog().lastIndex());
 238
 239             followerLogInformation.setMatchIndex(-1);
 240             followerLogInformation.setNextIndex(-1);
 241
 242             initiateCaptureSnapshot(followerId);
 243             updated = true;
 244         } else if (appendEntriesReply.isSuccess()) {
 245             updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 246         } else {
 247             LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
 248
 249             long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
 250             long followersLastLogTerm = getLogEntryTerm(followerLastLogIndex);
 251             if(appendEntriesReply.isForceInstallSnapshot()) {
 252                 // Reset the followers match and next index. This is to signal that this follower has nothing
 253                 // in common with this Leader and so would require a snapshot to be installed
 254                 followerLogInformation.setMatchIndex(-1);
 255                 followerLogInformation.setNextIndex(-1);
 256
 257                 // Force initiate a snapshot capture
 258                 initiateCaptureSnapshot(followerId);
 259             } else if(followerLastLogIndex < 0 || (followersLastLogTerm >= 0 &&
 260                     followersLastLogTerm == appendEntriesReply.getLogLastTerm())) {
 261                 // The follower's log is empty or the last entry is present in the leader's journal
 262                 // and the terms match so the follower is just behind the leader's journal from
 263                 // the last snapshot, if any. We'll catch up the follower quickly by starting at the
 264                 // follower's last log index.
 265
 266                 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 267             } else {
 268                 // The follower's log conflicts with leader's log so decrement follower's next index by 1
 269                 // in an attempt to find where the logs match.
 270
 271                 LOG.debug("{}: follower's last log term {} conflicts with the leader's {} - dec next index",
 272                         logName(), appendEntriesReply.getLogLastTerm(), followersLastLogTerm);
 273
 274                 followerLogInformation.decrNextIndex();
 275             }
 276         }
 277
 278         // Now figure out if this reply warrants a change in the commitIndex
 279         // If there exists an N such that N > commitIndex, a majority
 280         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 281         // set commitIndex = N (§5.3, §5.4).
 282         if(LOG.isTraceEnabled()) {
 283             LOG.trace("{}: handleAppendEntriesReply from {}: commitIndex: {}, lastAppliedIndex: {}, currentTerm: {}",
 284                     logName(), followerId, context.getCommitIndex(), context.getLastApplied(), currentTerm());
 285         }
 286
 287         for (long N = context.getCommitIndex() + 1; ; N++) {
 288             int replicatedCount = 1;
 289
 290             LOG.trace("{}: checking Nth index {}", logName(), N);
 291             for (FollowerLogInformation info : followerToLog.values()) {
 292                 final PeerInfo peerInfo = context.getPeerInfo(info.getId());
 293                 if(info.getMatchIndex() >= N && (peerInfo != null && peerInfo.isVoting())) {
 294                     replicatedCount++;
 295                 } else if(LOG.isTraceEnabled()) {
 296                     LOG.trace("{}: Not counting follower {} - matchIndex: {}, {}", logName(), info.getId(),
 297                             info.getMatchIndex(), peerInfo);
 298                 }
 299             }
 300
 301             if(LOG.isTraceEnabled()) {
 302                 LOG.trace("{}: replicatedCount {}, minReplicationCount: {}", logName(), replicatedCount, minReplicationCount);
 303             }
 304
 305             if (replicatedCount >= minReplicationCount) {
 306                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 307                 if (replicatedLogEntry == null) {
 308                     LOG.debug("{}: ReplicatedLogEntry not found for index {} - snapshotIndex: {}, journal size: {}",
 309                             logName(), N, context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().size());
 310                     break;
 311                 }
 312
 313                 // Don't update the commit index if the log entry is from a previous term, as per §5.4.1:
 314                 // "Raft never commits log entries from previous terms by counting replicas".
 315                 // However we keep looping so we can make progress when new entries in the current term
 316                 // reach consensus, as per §5.4.1: "once an entry from the current term is committed by
 317                 // counting replicas, then all prior entries are committed indirectly".
 318                 if (replicatedLogEntry.getTerm() == currentTerm()) {
 319                     LOG.trace("{}: Setting commit index to {}", logName(), N);
 320                     context.setCommitIndex(N);
 321                 } else {
 322                     LOG.debug("{}: Not updating commit index to {} - retrieved log entry with index {}, term {} does not match the current term {}",
 323                             logName(), N, replicatedLogEntry.getIndex(), replicatedLogEntry.getTerm(), currentTerm());
 324                 }
 325             } else {
 326                 LOG.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount);
 327                 break;
 328             }
 329         }
 330
 331         // Apply the change to the state machine
 332         if (context.getCommitIndex() > context.getLastApplied()) {
 333             if(LOG.isDebugEnabled()) {
 334                 LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
 335                         logName(), followerId, context.getCommitIndex(), context.getLastApplied());
 336             }
 337
 338             applyLogToStateMachine(context.getCommitIndex());
 339         }
 340
 341         if (!context.getSnapshotManager().isCapturing()) {
 342             purgeInMemoryLog();
 343         }
 344
 345         //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
 346         sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
 347
 348         return this;
 349     }
 350
 351     private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
 352             AppendEntriesReply appendEntriesReply) {
 353         boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
 354         updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
 355
 356         if(updated && LOG.isDebugEnabled()) {
 357             LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
 358                     logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
 359                     followerLogInformation.getNextIndex());
 360         }
 361         return updated;
 362     }
 363
 364     private void purgeInMemoryLog() {
 365         //find the lowest index across followers which has been replicated to all.
 366         // lastApplied if there are no followers, so that we keep clearing the log for single-node
 367         // we would delete the in-mem log from that index on, in-order to minimize mem usage
 368         // we would also share this info thru AE with the followers so that they can delete their log entries as well.
 369         long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
 370         for (FollowerLogInformation info : followerToLog.values()) {
 371             minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
 372         }
 373
 374         super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
 375     }
 376
 377     @Override
 378     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 379         final Iterator<ClientRequestTracker> it = trackers.iterator();
 380         while (it.hasNext()) {
 381             final ClientRequestTracker t = it.next();
 382             if (t.getIndex() == logIndex) {
 383                 it.remove();
 384                 return t;
 385             }
 386         }
 387
 388         return null;
 389     }
 390
 391     @Override
 392     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 393         RequestVoteReply requestVoteReply) {
 394         return this;
 395     }
 396
 397     protected void beforeSendHeartbeat(){}
 398
 399     @Override
 400     public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 401         Preconditions.checkNotNull(sender, "sender should not be null");
 402
 403         Object message = fromSerializableMessage(originalMessage);
 404
 405         if (message instanceof RaftRPC) {
 406             RaftRPC rpc = (RaftRPC) message;
 407             // If RPC request or response contains term T > currentTerm:
 408             // set currentTerm = T, convert to follower (§5.1)
 409             // This applies to all RPC messages and responses
 410             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 411                 LOG.debug("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
 412                         logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 413
 414                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 415
 416                 return internalSwitchBehavior(RaftState.Follower);
 417             }
 418         }
 419
 420         if (message instanceof SendHeartBeat) {
 421             beforeSendHeartbeat();
 422             sendHeartBeat();
 423             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 424         } else if(message instanceof SendInstallSnapshot) {
 425             // received from RaftActor
 426             setSnapshot(((SendInstallSnapshot) message).getSnapshot());
 427             sendInstallSnapshot();
 428         } else if (message instanceof Replicate) {
 429             replicate((Replicate) message);
 430         } else if (message instanceof InstallSnapshotReply) {
 431             handleInstallSnapshotReply((InstallSnapshotReply) message);
 432         } else {
 433             return super.handleMessage(sender, message);
 434         }
 435
 436         return this;
 437     }
 438
 439     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 440         LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
 441
 442         String followerId = reply.getFollowerId();
 443         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 444
 445         if (followerToSnapshot == null) {
 446             LOG.error("{}: FollowerToSnapshot not found for follower {} in InstallSnapshotReply",
 447                     logName(), followerId);
 448             return;
 449         }
 450
 451         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 452         if(followerLogInformation == null) {
 453             // This can happen during AddServer if it times out.
 454             LOG.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
 455                     logName(), followerId);
 456             mapFollowerToSnapshot.remove(followerId);
 457             return;
 458         }
 459
 460         followerLogInformation.markFollowerActive();
 461
 462         if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 463             boolean wasLastChunk = false;
 464             if (reply.isSuccess()) {
 465                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 466                     //this was the last chunk reply
 467                     if(LOG.isDebugEnabled()) {
 468                         LOG.debug("{}: InstallSnapshotReply received, " +
 469                                 "last chunk received, Chunk: {}. Follower: {} Setting nextIndex: {}",
 470                                 logName(), reply.getChunkIndex(), followerId,
 471                             context.getReplicatedLog().getSnapshotIndex() + 1
 472                         );
 473                     }
 474
 475                     long followerMatchIndex = snapshot.get().getLastIncludedIndex();
 476                     followerLogInformation.setMatchIndex(followerMatchIndex);
 477                     followerLogInformation.setNextIndex(followerMatchIndex + 1);
 478                     mapFollowerToSnapshot.remove(followerId);
 479
 480                     LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}",
 481                         logName(), followerId, followerLogInformation.getMatchIndex(),
 482                         followerLogInformation.getNextIndex());
 483
 484                     if (mapFollowerToSnapshot.isEmpty()) {
 485                         // once there are no pending followers receiving snapshots
 486                         // we can remove snapshot from the memory
 487                         setSnapshot(null);
 488                     }
 489                     wasLastChunk = true;
 490                     if(context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED){
 491                         UnInitializedFollowerSnapshotReply unInitFollowerSnapshotSuccess =
 492                                              new UnInitializedFollowerSnapshotReply(followerId);
 493                         context.getActor().tell(unInitFollowerSnapshotSuccess, context.getActor());
 494                         LOG.debug("Sent message UnInitializedFollowerSnapshotReply to self");
 495                     }
 496                 } else {
 497                     followerToSnapshot.markSendStatus(true);
 498                 }
 499             } else {
 500                 LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
 501                         logName(), reply.getChunkIndex());
 502
 503                 followerToSnapshot.markSendStatus(false);
 504             }
 505
 506             if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
 507                 // Since the follower is now caught up try to purge the log.
 508                 purgeInMemoryLog();
 509             } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
 510                 ActorSelection followerActor = context.getPeerActorSelection(followerId);
 511                 if(followerActor != null) {
 512                     sendSnapshotChunk(followerActor, followerId);
 513                 }
 514             }
 515
 516         } else {
 517             LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
 518                     logName(), reply.getChunkIndex(), followerId,
 519                     followerToSnapshot.getChunkIndex());
 520
 521             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 522                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 523                 // so that Installing the snapshot can resume from the beginning
 524                 followerToSnapshot.reset();
 525             }
 526         }
 527     }
 528
 529     private void replicate(Replicate replicate) {
 530         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 531
 532         LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}", logName(),
 533                 replicate.getIdentifier(), logIndex);
 534
 535         // Create a tracker entry we will use this later to notify the
 536         // client actor
 537         trackers.add(
 538             new ClientRequestTrackerImpl(replicate.getClientActor(),
 539                 replicate.getIdentifier(),
 540                 logIndex)
 541         );
 542
 543         boolean applyModificationToState = !context.anyVotingPeers()
 544                 || context.getRaftPolicy().applyModificationToStateBeforeConsensus();
 545
 546         if(applyModificationToState){
 547             context.setCommitIndex(logIndex);
 548             applyLogToStateMachine(logIndex);
 549         }
 550
 551         if (!followerToLog.isEmpty()) {
 552             sendAppendEntries(0, false);
 553         }
 554     }
 555
 556     protected void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) {
 557         // Send an AppendEntries to all followers
 558         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 559             final String followerId = e.getKey();
 560             final FollowerLogInformation followerLogInformation = e.getValue();
 561             // This checks helps not to send a repeat message to the follower
 562             if(!followerLogInformation.isFollowerActive() ||
 563                     followerLogInformation.timeSinceLastActivity() >= timeSinceLastActivityInterval) {
 564                 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
 565             }
 566         }
 567     }
 568
 569     /**
 570      *
 571      * This method checks if any update needs to be sent to the given follower. This includes append log entries,
 572      * sending next snapshot chunk, and initiating a snapshot.
 573      * @return true if any update is sent, false otherwise
 574      */
 575
 576     private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
 577                                        boolean sendHeartbeat, boolean isHeartbeat) {
 578
 579         ActorSelection followerActor = context.getPeerActorSelection(followerId);
 580         if (followerActor != null) {
 581             long followerNextIndex = followerLogInformation.getNextIndex();
 582             boolean isFollowerActive = followerLogInformation.isFollowerActive();
 583             boolean sendAppendEntries = false;
 584             List<ReplicatedLogEntry> entries = Collections.emptyList();
 585
 586             if (mapFollowerToSnapshot.get(followerId) != null) {
 587                 // if install snapshot is in process , then sent next chunk if possible
 588                 if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 589                     sendSnapshotChunk(followerActor, followerId);
 590                 } else if(sendHeartbeat) {
 591                     // we send a heartbeat even if we have not received a reply for the last chunk
 592                     sendAppendEntries = true;
 593                 }
 594             } else {
 595                 long leaderLastIndex = context.getReplicatedLog().lastIndex();
 596                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 597
 598                 if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
 599                     LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
 600                             logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
 601                 }
 602
 603                 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
 604
 605                     LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
 606                             followerNextIndex, followerId);
 607
 608                     if(followerLogInformation.okToReplicate()) {
 609                         // Try to send all the entries in the journal but not exceeding the max data size
 610                         // for a single AppendEntries message.
 611                         int maxEntries = (int) context.getReplicatedLog().size();
 612                         entries = context.getReplicatedLog().getFrom(followerNextIndex, maxEntries,
 613                                 context.getConfigParams().getSnapshotChunkSize());
 614                         sendAppendEntries = true;
 615                     }
 616                 } else if (isFollowerActive && followerNextIndex >= 0 &&
 617                     leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
 618                     // if the followers next index is not present in the leaders log, and
 619                     // if the follower is just not starting and if leader's index is more than followers index
 620                     // then snapshot should be sent
 621
 622                     if (LOG.isDebugEnabled()) {
 623                         LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s, " +
 624                                     "follower-nextIndex: %d, leader-snapshot-index: %d,  " +
 625                                     "leader-last-index: %d", logName(), followerId,
 626                                     followerNextIndex, leaderSnapShotIndex, leaderLastIndex));
 627                     }
 628
 629                     // Send heartbeat to follower whenever install snapshot is initiated.
 630                     sendAppendEntries = true;
 631                     if (canInstallSnapshot(followerNextIndex)) {
 632                         initiateCaptureSnapshot(followerId);
 633                     }
 634
 635                 } else if(sendHeartbeat) {
 636                     // we send an AppendEntries, even if the follower is inactive
 637                     // in-order to update the followers timestamp, in case it becomes active again
 638                     sendAppendEntries = true;
 639                 }
 640
 641             }
 642
 643             if(sendAppendEntries) {
 644                 sendAppendEntriesToFollower(followerActor, followerNextIndex,
 645                         entries, followerId);
 646             }
 647         }
 648     }
 649
 650     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 651         List<ReplicatedLogEntry> entries, String followerId) {
 652         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
 653             getLogEntryIndex(followerNextIndex - 1),
 654             getLogEntryTerm(followerNextIndex - 1), entries,
 655             context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
 656
 657         if(!entries.isEmpty() || LOG.isTraceEnabled()) {
 658             LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
 659                     appendEntries);
 660         }
 661
 662         followerActor.tell(appendEntries, actor());
 663     }
 664
 665     /**
 666      * Install Snapshot works as follows
 667      * 1. Leader initiates the capture snapshot by sending a CaptureSnapshot message to actor
 668      * 2. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 669      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 670      * 3. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 671      * 4. On complete, Follower sends back a InstallSnapshotReply.
 672      * 5. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 673      * and replenishes the memory by deleting the snapshot in Replicated log.
 674      * 6. If another follower requires a snapshot and a snapshot has been collected (via CaptureSnapshotReply)
 675      * then send the existing snapshot in chunks to the follower.
 676      * @param followerId
 677      */
 678     public boolean initiateCaptureSnapshot(String followerId) {
 679         if (snapshot.isPresent()) {
 680             // if a snapshot is present in the memory, most likely another install is in progress
 681             // no need to capture snapshot.
 682             // This could happen if another follower needs an install when one is going on.
 683             final ActorSelection followerActor = context.getPeerActorSelection(followerId);
 684             sendSnapshotChunk(followerActor, followerId);
 685             return true;
 686         } else {
 687             return context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
 688                     this.getReplicatedToAllIndex(), followerId);
 689         }
 690     }
 691
 692     private boolean canInstallSnapshot(long nextIndex){
 693         // If the follower's nextIndex is -1 then we might as well send it a snapshot
 694         // Otherwise send it a snapshot only if the nextIndex is not present in the log but is present
 695         // in the snapshot
 696         return (nextIndex == -1 ||
 697                 (!context.getReplicatedLog().isPresent(nextIndex)
 698                         && context.getReplicatedLog().isInSnapshot(nextIndex)));
 699
 700     }
 701
 702
 703     private void sendInstallSnapshot() {
 704         LOG.debug("{}: sendInstallSnapshot", logName());
 705         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 706             String followerId = e.getKey();
 707             ActorSelection followerActor = context.getPeerActorSelection(followerId);
 708             FollowerLogInformation followerLogInfo = e.getValue();
 709
 710             if (followerActor != null) {
 711                 long nextIndex = followerLogInfo.getNextIndex();
 712                 if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED ||
 713                         canInstallSnapshot(nextIndex)) {
 714                     sendSnapshotChunk(followerActor, followerId);
 715                 }
 716             }
 717         }
 718     }
 719
 720     /**
 721      *  Sends a snapshot chunk to a given follower
 722      *  InstallSnapshot should qualify as a heartbeat too.
 723      */
 724     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 725         try {
 726             if (snapshot.isPresent()) {
 727                 byte[] nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes());
 728
 729                 // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
 730                 // followerId to the followerToSnapshot map.
 731                 FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 732
 733                 int nextChunkIndex = followerToSnapshot.incrementChunkIndex();
 734                 Optional<ServerConfigurationPayload> serverConfig = Optional.absent();
 735                 if(followerToSnapshot.isLastChunk(nextChunkIndex)) {
 736                     serverConfig = Optional.fromNullable(context.getPeerServerInfo(true));
 737                 }
 738
 739                 followerActor.tell(
 740                     new InstallSnapshot(currentTerm(), context.getId(),
 741                         snapshot.get().getLastIncludedIndex(),
 742                         snapshot.get().getLastIncludedTerm(),
 743                         nextSnapshotChunk,
 744                         nextChunkIndex,
 745                         followerToSnapshot.getTotalChunks(),
 746                         Optional.of(followerToSnapshot.getLastChunkHashCode()),
 747                         serverConfig
 748                     ).toSerializable(followerToLog.get(followerId).getRaftVersion()),
 749                     actor()
 750                 );
 751
 752                 if(LOG.isDebugEnabled()) {
 753                     LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
 754                             logName(), followerActor.path(), followerToSnapshot.getChunkIndex(),
 755                             followerToSnapshot.getTotalChunks());
 756                 }
 757             }
 758         } catch (IOException e) {
 759             LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
 760         }
 761     }
 762
 763     /**
 764      * Acccepts snaphot as ByteString, enters into map for future chunks
 765      * creates and return a ByteString chunk
 766      */
 767     private byte[] getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 768         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 769         if (followerToSnapshot == null) {
 770             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 771             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 772         }
 773         byte[] nextChunk = followerToSnapshot.getNextChunk();
 774
 775         LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.length);
 776
 777         return nextChunk;
 778     }
 779
 780     private void sendHeartBeat() {
 781         if (!followerToLog.isEmpty()) {
 782             LOG.trace("{}: Sending heartbeat", logName());
 783             sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toMillis(), true);
 784         }
 785     }
 786
 787     private void stopHeartBeat() {
 788         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 789             heartbeatSchedule.cancel();
 790         }
 791     }
 792
 793     private void scheduleHeartBeat(FiniteDuration interval) {
 794         if (followerToLog.isEmpty()) {
 795             // Optimization - do not bother scheduling a heartbeat as there are
 796             // no followers
 797             return;
 798         }
 799
 800         stopHeartBeat();
 801
 802         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 803         // message is sent to itself.
 804         // Scheduling the heartbeat only once here because heartbeats do not
 805         // need to be sent if there are other messages being sent to the remote
 806         // actor.
 807         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 808             interval, context.getActor(), SendHeartBeat.INSTANCE,
 809             context.getActorSystem().dispatcher(), context.getActor());
 810     }
 811
 812     @Override
 813     public void close() {
 814         stopHeartBeat();
 815     }
 816
 817     @Override
 818     public final String getLeaderId() {
 819         return context.getId();
 820     }
 821
 822     @Override
 823     public final short getLeaderPayloadVersion() {
 824         return context.getPayloadVersion();
 825     }
 826
 827     protected boolean isLeaderIsolated() {
 828         int minPresent = getMinIsolatedLeaderPeerCount();
 829         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 830             final PeerInfo peerInfo = context.getPeerInfo(followerLogInformation.getId());
 831             if(peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) {
 832                 --minPresent;
 833                 if (minPresent == 0) {
 834                     return false;
 835                 }
 836             }
 837         }
 838         return (minPresent != 0);
 839     }
 840
 841     /**
 842      * Encapsulates the snapshot bytestring and handles the logic of sending
 843      * snapshot chunks
 844      */
 845     protected class FollowerToSnapshot {
 846         private final ByteString snapshotBytes;
 847         private int offset = 0;
 848         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 849         private int replyReceivedForOffset;
 850         // if replyStatus is false, the previous chunk is attempted
 851         private boolean replyStatus = false;
 852         private int chunkIndex;
 853         private final int totalChunks;
 854         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 855         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 856
 857         public FollowerToSnapshot(ByteString snapshotBytes) {
 858             this.snapshotBytes = snapshotBytes;
 859             int size = snapshotBytes.size();
 860             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 861                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 862             if(LOG.isDebugEnabled()) {
 863                 LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}",
 864                         logName(), size, totalChunks);
 865             }
 866             replyReceivedForOffset = -1;
 867             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 868         }
 869
 870         public ByteString getSnapshotBytes() {
 871             return snapshotBytes;
 872         }
 873
 874         public int incrementOffset() {
 875             if(replyStatus) {
 876                 // if prev chunk failed, we would want to sent the same chunk again
 877                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 878             }
 879             return offset;
 880         }
 881
 882         public int incrementChunkIndex() {
 883             if (replyStatus) {
 884                 // if prev chunk failed, we would want to sent the same chunk again
 885                 chunkIndex =  chunkIndex + 1;
 886             }
 887             return chunkIndex;
 888         }
 889
 890         public int getChunkIndex() {
 891             return chunkIndex;
 892         }
 893
 894         public int getTotalChunks() {
 895             return totalChunks;
 896         }
 897
 898         public boolean canSendNextChunk() {
 899             // we only send a false if a chunk is sent but we have not received a reply yet
 900             return replyReceivedForOffset == offset;
 901         }
 902
 903         public boolean isLastChunk(int chunkIndex) {
 904             return totalChunks == chunkIndex;
 905         }
 906
 907         public void markSendStatus(boolean success) {
 908             if (success) {
 909                 // if the chunk sent was successful
 910                 replyReceivedForOffset = offset;
 911                 replyStatus = true;
 912                 lastChunkHashCode = nextChunkHashCode;
 913             } else {
 914                 // if the chunk sent was failure
 915                 replyReceivedForOffset = offset;
 916                 replyStatus = false;
 917             }
 918         }
 919
 920         public byte[] getNextChunk() {
 921             int snapshotLength = getSnapshotBytes().size();
 922             int start = incrementOffset();
 923             int size = context.getConfigParams().getSnapshotChunkSize();
 924             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 925                 size = snapshotLength;
 926             } else if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 927                 size = snapshotLength - start;
 928             }
 929
 930             byte[] nextChunk = new byte[size];
 931             getSnapshotBytes().copyTo(nextChunk, start, 0, size);
 932             nextChunkHashCode = Arrays.hashCode(nextChunk);
 933
 934             LOG.debug("{}: Next chunk: total length={}, offset={}, size={}, hashCode={}", logName(),
 935                     snapshotLength, start, size, nextChunkHashCode);
 936             return nextChunk;
 937         }
 938
 939         /**
 940          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 941          */
 942         public void reset(){
 943             offset = 0;
 944             replyStatus = false;
 945             replyReceivedForOffset = offset;
 946             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 947             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 948         }
 949
 950         public int getLastChunkHashCode() {
 951             return lastChunkHashCode;
 952         }
 953     }
 954
 955     // called from example-actor for printing the follower-states
 956     public String printFollowerStates() {
 957         final StringBuilder sb = new StringBuilder();
 958
 959         sb.append('[');
 960         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 961             sb.append('{');
 962             sb.append(followerLogInformation.getId());
 963             sb.append(" state:");
 964             sb.append(followerLogInformation.isFollowerActive());
 965             sb.append("},");
 966         }
 967         sb.append(']');
 968
 969         return sb.toString();
 970     }
 971
 972     @VisibleForTesting
 973     public FollowerLogInformation getFollower(String followerId) {
 974         return followerToLog.get(followerId);
 975     }
 976
 977     @VisibleForTesting
 978     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 979         mapFollowerToSnapshot.put(followerId, snapshot);
 980     }
 981
 982     @VisibleForTesting
 983     public int followerSnapshotSize() {
 984         return mapFollowerToSnapshot.size();
 985     }
 986
 987     @VisibleForTesting
 988     public int followerLogSize() {
 989         return followerToLog.size();
 990     }
 991
 992     private static class SnapshotHolder {
 993         private final long lastIncludedTerm;
 994         private final long lastIncludedIndex;
 995         private final ByteString snapshotBytes;
 996
 997         SnapshotHolder(Snapshot snapshot) {
 998             this.lastIncludedTerm = snapshot.getLastAppliedTerm();
 999             this.lastIncludedIndex = snapshot.getLastAppliedIndex();
1000             this.snapshotBytes = ByteString.copyFrom(snapshot.getState());
1001         }
1002
1003         long getLastIncludedTerm() {
1004             return lastIncludedTerm;
1005         }
1006
1007         long getLastIncludedIndex() {
1008             return lastIncludedIndex;
1009         }
1010
1011         ByteString getSnapshotBytes() {
1012             return snapshotBytes;
1013         }
1014     }
1015 }