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