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.Snapshot;
  39 import org.opendaylight.controller.cluster.raft.VotingState;
  40 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  41 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
  42 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
  43 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  44 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  45 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  46 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  47 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  48 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  49 import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply;
  50 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
  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 = Optional.absent();;
 100     private int minReplicationCount;
 101
 102     protected AbstractLeader(RaftActorContext context, RaftState state,
 103             @Nullable AbstractLeader initializeFromLeader) {
 104         super(context, state);
 105
 106         if(initializeFromLeader != null) {
 107             followerToLog.putAll(initializeFromLeader.followerToLog);
 108             mapFollowerToSnapshot.putAll(initializeFromLeader.mapFollowerToSnapshot);
 109             snapshot = initializeFromLeader.snapshot;
 110             trackers.addAll(initializeFromLeader.trackers);
 111         } else {
 112             for(PeerInfo peerInfo: context.getPeers()) {
 113                 FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(peerInfo, -1, context);
 114                 followerToLog.put(peerInfo.getId(), followerLogInformation);
 115             }
 116         }
 117
 118         LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
 119
 120         updateMinReplicaCount();
 121
 122         // Immediately schedule a heartbeat
 123         // Upon election: send initial empty AppendEntries RPCs
 124         // (heartbeat) to each server; repeat during idle periods to
 125         // prevent election timeouts (§5.2)
 126         sendAppendEntries(0, false);
 127
 128         // It is important to schedule this heartbeat here
 129         scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 130     }
 131
 132     protected AbstractLeader(RaftActorContext context, RaftState state) {
 133         this(context, state, null);
 134     }
 135
 136     /**
 137      * Return an immutable collection of follower identifiers.
 138      *
 139      * @return Collection of follower IDs
 140      */
 141     public final Collection<String> getFollowerIds() {
 142         return followerToLog.keySet();
 143     }
 144
 145     public void addFollower(String followerId) {
 146         FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(
 147                 context.getPeerInfo(followerId), -1, context);
 148         followerToLog.put(followerId, followerLogInformation);
 149
 150         if(heartbeatSchedule == null) {
 151             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 152         }
 153     }
 154
 155     public void removeFollower(String followerId) {
 156         followerToLog.remove(followerId);
 157         mapFollowerToSnapshot.remove(followerId);
 158     }
 159
 160     public void updateMinReplicaCount() {
 161         int numVoting = 0;
 162         for(PeerInfo peer: context.getPeers()) {
 163             if(peer.isVoting()) {
 164                 numVoting++;
 165             }
 166         }
 167
 168         minReplicationCount = getMajorityVoteCount(numVoting);
 169     }
 170
 171     protected int getMinIsolatedLeaderPeerCount(){
 172       //the isolated Leader peer count will be 1 less than the majority vote count.
 173         //this is because the vote count has the self vote counted in it
 174         //for e.g
 175         //0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
 176         //2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
 177         //4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
 178
 179         return minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
 180     }
 181
 182     @VisibleForTesting
 183     void setSnapshot(@Nullable Snapshot snapshot) {
 184         if(snapshot != null) {
 185             this.snapshot = Optional.of(new SnapshotHolder(snapshot));
 186         } else {
 187             this.snapshot = Optional.absent();
 188         }
 189     }
 190
 191     @Override
 192     protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 193         AppendEntries appendEntries) {
 194
 195         LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 196
 197         return this;
 198     }
 199
 200     @Override
 201     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 202         AppendEntriesReply appendEntriesReply) {
 203
 204         if(LOG.isTraceEnabled()) {
 205             LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
 206         }
 207
 208         // Update the FollowerLogInformation
 209         String followerId = appendEntriesReply.getFollowerId();
 210         FollowerLogInformation followerLogInformation =
 211             followerToLog.get(followerId);
 212
 213         if(followerLogInformation == null){
 214             LOG.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
 215             return this;
 216         }
 217
 218         if(followerLogInformation.timeSinceLastActivity() >
 219                 context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
 220             LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
 221                             "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
 222                     logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
 223                     context.getLastApplied(), context.getCommitIndex());
 224         }
 225
 226         followerLogInformation.markFollowerActive();
 227         followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
 228         followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
 229
 230         boolean updated = false;
 231         if(appendEntriesReply.getLogLastIndex() > context.getReplicatedLog().lastIndex()) {
 232             // The follower's log is actually ahead of the leader's log. Normally this doesn't happen
 233             // in raft as a node cannot become leader if it's log is behind another's. However, the
 234             // non-voting semantics deviate a bit from raft. Only voting members participate in
 235             // elections and can become leader so it's possible for a non-voting follower to be ahead
 236             // of the leader. This can happen if persistence is disabled and all voting members are
 237             // restarted. In this case, the voting leader will start out with an empty log however
 238             // the non-voting followers still retain the previous data in memory. On the first
 239             // AppendEntries, the non-voting follower returns a successful reply b/c the prevLogIndex
 240             // sent by the leader is -1 and thus the integrity checks pass. However the follower's returned
 241             // lastLogIndex may be higher in which case we want to reset the follower by installing a
 242             // snapshot. It's also possible that the follower's last log index is behind the leader's.
 243             // However in this case the log terms won't match and the logs will conflict - this is handled
 244             // elsewhere.
 245             LOG.debug("{}: handleAppendEntriesReply: follower {} lastIndex {} is ahead of our lastIndex {} - forcing install snaphot",
 246                     logName(), followerLogInformation.getId(), appendEntriesReply.getLogLastIndex(),
 247                     context.getReplicatedLog().lastIndex());
 248
 249             followerLogInformation.setMatchIndex(-1);
 250             followerLogInformation.setNextIndex(-1);
 251
 252             initiateCaptureSnapshot(followerId);
 253             updated = true;
 254         } else if (appendEntriesReply.isSuccess()) {
 255             updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 256         } else {
 257             LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
 258
 259             long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
 260             long followersLastLogTerm = getLogEntryTerm(followerLastLogIndex);
 261             if(appendEntriesReply.isForceInstallSnapshot()) {
 262                 // Reset the followers match and next index. This is to signal that this follower has nothing
 263                 // in common with this Leader and so would require a snapshot to be installed
 264                 followerLogInformation.setMatchIndex(-1);
 265                 followerLogInformation.setNextIndex(-1);
 266
 267                 // Force initiate a snapshot capture
 268                 initiateCaptureSnapshot(followerId);
 269             } else if(followerLastLogIndex < 0 || (followersLastLogTerm >= 0 &&
 270                     followersLastLogTerm == appendEntriesReply.getLogLastTerm())) {
 271                 // The follower's log is empty or the last entry is present in the leader's journal
 272                 // and the terms match so the follower is just behind the leader's journal from
 273                 // the last snapshot, if any. We'll catch up the follower quickly by starting at the
 274                 // follower's last log index.
 275
 276                 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
 277             } else {
 278                 // The follower's log conflicts with leader's log so decrement follower's next index by 1
 279                 // in an attempt to find where the logs match.
 280
 281                 LOG.debug("{}: follower's last log term {} conflicts with the leader's {} - dec next index",
 282                         logName(), appendEntriesReply.getLogLastTerm(), followersLastLogTerm);
 283
 284                 followerLogInformation.decrNextIndex();
 285             }
 286         }
 287
 288         // Now figure out if this reply warrants a change in the commitIndex
 289         // If there exists an N such that N > commitIndex, a majority
 290         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 291         // set commitIndex = N (§5.3, §5.4).
 292         if(LOG.isTraceEnabled()) {
 293             LOG.trace("{}: handleAppendEntriesReply from {}: commitIndex: {}, lastAppliedIndex: {}, currentTerm: {}",
 294                     logName(), followerId, context.getCommitIndex(), context.getLastApplied(), currentTerm());
 295         }
 296
 297         for (long N = context.getCommitIndex() + 1; ; N++) {
 298             int replicatedCount = 1;
 299
 300             LOG.trace("{}: checking Nth index {}", logName(), N);
 301             for (FollowerLogInformation info : followerToLog.values()) {
 302                 final PeerInfo peerInfo = context.getPeerInfo(info.getId());
 303                 if(info.getMatchIndex() >= N && peerInfo != null && peerInfo.isVoting()) {
 304                     replicatedCount++;
 305                 } else if(LOG.isTraceEnabled()) {
 306                     LOG.trace("{}: Not counting follower {} - matchIndex: {}, {}", logName(), info.getId(),
 307                             info.getMatchIndex(), peerInfo);
 308                 }
 309             }
 310
 311             if(LOG.isTraceEnabled()) {
 312                 LOG.trace("{}: replicatedCount {}, minReplicationCount: {}", logName(), replicatedCount, minReplicationCount);
 313             }
 314
 315             if (replicatedCount >= minReplicationCount) {
 316                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
 317                 if (replicatedLogEntry == null) {
 318                     LOG.debug("{}: ReplicatedLogEntry not found for index {} - snapshotIndex: {}, journal size: {}",
 319                             logName(), N, context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().size());
 320                     break;
 321                 }
 322
 323                 // Don't update the commit index if the log entry is from a previous term, as per §5.4.1:
 324                 // "Raft never commits log entries from previous terms by counting replicas".
 325                 // However we keep looping so we can make progress when new entries in the current term
 326                 // reach consensus, as per §5.4.1: "once an entry from the current term is committed by
 327                 // counting replicas, then all prior entries are committed indirectly".
 328                 if (replicatedLogEntry.getTerm() == currentTerm()) {
 329                     LOG.trace("{}: Setting commit index to {}", logName(), N);
 330                     context.setCommitIndex(N);
 331                 } else {
 332                     LOG.debug("{}: Not updating commit index to {} - retrieved log entry with index {}, term {} does not match the current term {}",
 333                             logName(), N, replicatedLogEntry.getIndex(), replicatedLogEntry.getTerm(), currentTerm());
 334                 }
 335             } else {
 336                 LOG.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount);
 337                 break;
 338             }
 339         }
 340
 341         // Apply the change to the state machine
 342         if (context.getCommitIndex() > context.getLastApplied()) {
 343             if(LOG.isDebugEnabled()) {
 344                 LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
 345                         logName(), followerId, context.getCommitIndex(), context.getLastApplied());
 346             }
 347
 348             applyLogToStateMachine(context.getCommitIndex());
 349         }
 350
 351         if (!context.getSnapshotManager().isCapturing()) {
 352             purgeInMemoryLog();
 353         }
 354
 355         //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
 356         sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
 357
 358         return this;
 359     }
 360
 361     private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
 362             AppendEntriesReply appendEntriesReply) {
 363         boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
 364         updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
 365
 366         if(updated && LOG.isDebugEnabled()) {
 367             LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
 368                     logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
 369                     followerLogInformation.getNextIndex());
 370         }
 371         return updated;
 372     }
 373
 374     private void purgeInMemoryLog() {
 375         //find the lowest index across followers which has been replicated to all.
 376         // lastApplied if there are no followers, so that we keep clearing the log for single-node
 377         // we would delete the in-mem log from that index on, in-order to minimize mem usage
 378         // we would also share this info thru AE with the followers so that they can delete their log entries as well.
 379         long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
 380         for (FollowerLogInformation info : followerToLog.values()) {
 381             minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
 382         }
 383
 384         super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
 385     }
 386
 387     @Override
 388     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 389         final Iterator<ClientRequestTracker> it = trackers.iterator();
 390         while (it.hasNext()) {
 391             final ClientRequestTracker t = it.next();
 392             if (t.getIndex() == logIndex) {
 393                 it.remove();
 394                 return t;
 395             }
 396         }
 397
 398         return null;
 399     }
 400
 401     @Override
 402     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 403         RequestVoteReply requestVoteReply) {
 404         return this;
 405     }
 406
 407     protected void beforeSendHeartbeat(){}
 408
 409     @Override
 410     public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
 411         Preconditions.checkNotNull(sender, "sender should not be null");
 412
 413         if (message instanceof RaftRPC) {
 414             RaftRPC rpc = (RaftRPC) message;
 415             // If RPC request or response contains term T > currentTerm:
 416             // set currentTerm = T, convert to follower (§5.1)
 417             // This applies to all RPC messages and responses
 418             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 419                 LOG.debug("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
 420                         logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 421
 422                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 423
 424                 return internalSwitchBehavior(RaftState.Follower);
 425             }
 426         }
 427
 428         if (message instanceof SendHeartBeat) {
 429             beforeSendHeartbeat();
 430             sendHeartBeat();
 431             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 432         } else if(message instanceof SendInstallSnapshot) {
 433             // received from RaftActor
 434             setSnapshot(((SendInstallSnapshot) message).getSnapshot());
 435             sendInstallSnapshot();
 436         } else if (message instanceof Replicate) {
 437             replicate((Replicate) message);
 438         } else if (message instanceof InstallSnapshotReply) {
 439             handleInstallSnapshotReply((InstallSnapshotReply) message);
 440         } else {
 441             return super.handleMessage(sender, message);
 442         }
 443
 444         return this;
 445     }
 446
 447     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 448         LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
 449
 450         String followerId = reply.getFollowerId();
 451         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 452
 453         if (followerToSnapshot == null) {
 454             LOG.error("{}: FollowerToSnapshot not found for follower {} in InstallSnapshotReply",
 455                     logName(), followerId);
 456             return;
 457         }
 458
 459         FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 460         if(followerLogInformation == null) {
 461             // This can happen during AddServer if it times out.
 462             LOG.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
 463                     logName(), followerId);
 464             mapFollowerToSnapshot.remove(followerId);
 465             return;
 466         }
 467
 468         followerLogInformation.markFollowerActive();
 469
 470         if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 471             boolean wasLastChunk = false;
 472             if (reply.isSuccess()) {
 473                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 474                     //this was the last chunk reply
 475                     if(LOG.isDebugEnabled()) {
 476                         LOG.debug("{}: InstallSnapshotReply received, " +
 477                                 "last chunk received, Chunk: {}. Follower: {} Setting nextIndex: {}",
 478                                 logName(), reply.getChunkIndex(), followerId,
 479                             context.getReplicatedLog().getSnapshotIndex() + 1
 480                         );
 481                     }
 482
 483                     long followerMatchIndex = snapshot.get().getLastIncludedIndex();
 484                     followerLogInformation.setMatchIndex(followerMatchIndex);
 485                     followerLogInformation.setNextIndex(followerMatchIndex + 1);
 486                     mapFollowerToSnapshot.remove(followerId);
 487
 488                     LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}",
 489                         logName(), followerId, followerLogInformation.getMatchIndex(),
 490                         followerLogInformation.getNextIndex());
 491
 492                     if (mapFollowerToSnapshot.isEmpty()) {
 493                         // once there are no pending followers receiving snapshots
 494                         // we can remove snapshot from the memory
 495                         setSnapshot(null);
 496                     }
 497                     wasLastChunk = true;
 498                     if(context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED){
 499                         UnInitializedFollowerSnapshotReply unInitFollowerSnapshotSuccess =
 500                                              new UnInitializedFollowerSnapshotReply(followerId);
 501                         context.getActor().tell(unInitFollowerSnapshotSuccess, context.getActor());
 502                         LOG.debug("Sent message UnInitializedFollowerSnapshotReply to self");
 503                     }
 504                 } else {
 505                     followerToSnapshot.markSendStatus(true);
 506                 }
 507             } else {
 508                 LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
 509                         logName(), reply.getChunkIndex());
 510
 511                 followerToSnapshot.markSendStatus(false);
 512             }
 513
 514             if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
 515                 // Since the follower is now caught up try to purge the log.
 516                 purgeInMemoryLog();
 517             } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
 518                 ActorSelection followerActor = context.getPeerActorSelection(followerId);
 519                 if(followerActor != null) {
 520                     sendSnapshotChunk(followerActor, followerId);
 521                 }
 522             }
 523
 524         } else {
 525             LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
 526                     logName(), reply.getChunkIndex(), followerId,
 527                     followerToSnapshot.getChunkIndex());
 528
 529             if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){
 530                 // Since the Follower did not find this index to be valid we should reset the follower snapshot
 531                 // so that Installing the snapshot can resume from the beginning
 532                 followerToSnapshot.reset();
 533             }
 534         }
 535     }
 536
 537     private void replicate(Replicate replicate) {
 538         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 539
 540         LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}", logName(),
 541                 replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass());
 542
 543         // Create a tracker entry we will use this later to notify the
 544         // client actor
 545         if(replicate.getClientActor() != null) {
 546             trackers.add(new ClientRequestTrackerImpl(replicate.getClientActor(), replicate.getIdentifier(),
 547                     logIndex));
 548         }
 549
 550         boolean applyModificationToState = !context.anyVotingPeers()
 551                 || context.getRaftPolicy().applyModificationToStateBeforeConsensus();
 552
 553         if(applyModificationToState){
 554             context.setCommitIndex(logIndex);
 555             applyLogToStateMachine(logIndex);
 556         }
 557
 558         if (!followerToLog.isEmpty()) {
 559             sendAppendEntries(0, false);
 560         }
 561     }
 562
 563     protected void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) {
 564         // Send an AppendEntries to all followers
 565         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 566             final String followerId = e.getKey();
 567             final FollowerLogInformation followerLogInformation = e.getValue();
 568             // This checks helps not to send a repeat message to the follower
 569             if(!followerLogInformation.isFollowerActive() ||
 570                     followerLogInformation.timeSinceLastActivity() >= timeSinceLastActivityInterval) {
 571                 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
 572             }
 573         }
 574     }
 575
 576     /**
 577      *
 578      * This method checks if any update needs to be sent to the given follower. This includes append log entries,
 579      * sending next snapshot chunk, and initiating a snapshot.
 580      * @return true if any update is sent, false otherwise
 581      */
 582
 583     private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
 584                                        boolean sendHeartbeat, boolean isHeartbeat) {
 585
 586         ActorSelection followerActor = context.getPeerActorSelection(followerId);
 587         if (followerActor != null) {
 588             long followerNextIndex = followerLogInformation.getNextIndex();
 589             boolean isFollowerActive = followerLogInformation.isFollowerActive();
 590             boolean sendAppendEntries = false;
 591             List<ReplicatedLogEntry> entries = Collections.emptyList();
 592
 593             if (mapFollowerToSnapshot.get(followerId) != null) {
 594                 // if install snapshot is in process , then sent next chunk if possible
 595                 if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 596                     sendSnapshotChunk(followerActor, followerId);
 597                 } else if(sendHeartbeat) {
 598                     // we send a heartbeat even if we have not received a reply for the last chunk
 599                     sendAppendEntries = true;
 600                 }
 601             } else {
 602                 long leaderLastIndex = context.getReplicatedLog().lastIndex();
 603                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 604
 605                 if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
 606                     LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
 607                             logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
 608                 }
 609
 610                 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
 611
 612                     LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
 613                             followerNextIndex, followerId);
 614
 615                     if(followerLogInformation.okToReplicate()) {
 616                         // Try to send all the entries in the journal but not exceeding the max data size
 617                         // for a single AppendEntries message.
 618                         int maxEntries = (int) context.getReplicatedLog().size();
 619                         entries = context.getReplicatedLog().getFrom(followerNextIndex, maxEntries,
 620                                 context.getConfigParams().getSnapshotChunkSize());
 621                         sendAppendEntries = true;
 622                     }
 623                 } else if (isFollowerActive && followerNextIndex >= 0 &&
 624                     leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
 625                     // if the followers next index is not present in the leaders log, and
 626                     // if the follower is just not starting and if leader's index is more than followers index
 627                     // then snapshot should be sent
 628
 629                     if (LOG.isDebugEnabled()) {
 630                         LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s, " +
 631                                     "follower-nextIndex: %d, leader-snapshot-index: %d,  " +
 632                                     "leader-last-index: %d", logName(), followerId,
 633                                     followerNextIndex, leaderSnapShotIndex, leaderLastIndex));
 634                     }
 635
 636                     // Send heartbeat to follower whenever install snapshot is initiated.
 637                     sendAppendEntries = true;
 638                     if (canInstallSnapshot(followerNextIndex)) {
 639                         initiateCaptureSnapshot(followerId);
 640                     }
 641
 642                 } else if(sendHeartbeat) {
 643                     // we send an AppendEntries, even if the follower is inactive
 644                     // in-order to update the followers timestamp, in case it becomes active again
 645                     sendAppendEntries = true;
 646                 }
 647
 648             }
 649
 650             if(sendAppendEntries) {
 651                 sendAppendEntriesToFollower(followerActor, followerNextIndex,
 652                         entries, followerId);
 653             }
 654         }
 655     }
 656
 657     private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
 658         List<ReplicatedLogEntry> entries, String followerId) {
 659         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
 660             getLogEntryIndex(followerNextIndex - 1),
 661             getLogEntryTerm(followerNextIndex - 1), entries,
 662             context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
 663
 664         if(!entries.isEmpty() || LOG.isTraceEnabled()) {
 665             LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
 666                     appendEntries);
 667         }
 668
 669         followerActor.tell(appendEntries, actor());
 670     }
 671
 672     /**
 673      * Install Snapshot works as follows
 674      * 1. Leader initiates the capture snapshot by sending a CaptureSnapshot message to actor
 675      * 2. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
 676      * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
 677      * 3. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
 678      * 4. On complete, Follower sends back a InstallSnapshotReply.
 679      * 5. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
 680      * and replenishes the memory by deleting the snapshot in Replicated log.
 681      * 6. If another follower requires a snapshot and a snapshot has been collected (via CaptureSnapshotReply)
 682      * then send the existing snapshot in chunks to the follower.
 683      * @param followerId
 684      */
 685     public boolean initiateCaptureSnapshot(String followerId) {
 686         if (snapshot.isPresent()) {
 687             // if a snapshot is present in the memory, most likely another install is in progress
 688             // no need to capture snapshot.
 689             // This could happen if another follower needs an install when one is going on.
 690             final ActorSelection followerActor = context.getPeerActorSelection(followerId);
 691             sendSnapshotChunk(followerActor, followerId);
 692             return true;
 693         } else {
 694             return context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
 695                     this.getReplicatedToAllIndex(), followerId);
 696         }
 697     }
 698
 699     private boolean canInstallSnapshot(long nextIndex){
 700         // If the follower's nextIndex is -1 then we might as well send it a snapshot
 701         // Otherwise send it a snapshot only if the nextIndex is not present in the log but is present
 702         // in the snapshot
 703         return nextIndex == -1 ||
 704                 (!context.getReplicatedLog().isPresent(nextIndex)
 705                         && context.getReplicatedLog().isInSnapshot(nextIndex));
 706
 707     }
 708
 709
 710     private void sendInstallSnapshot() {
 711         LOG.debug("{}: sendInstallSnapshot", logName());
 712         for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
 713             String followerId = e.getKey();
 714             ActorSelection followerActor = context.getPeerActorSelection(followerId);
 715             FollowerLogInformation followerLogInfo = e.getValue();
 716
 717             if (followerActor != null) {
 718                 long nextIndex = followerLogInfo.getNextIndex();
 719                 if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED ||
 720                         canInstallSnapshot(nextIndex)) {
 721                     sendSnapshotChunk(followerActor, followerId);
 722                 }
 723             }
 724         }
 725     }
 726
 727     /**
 728      *  Sends a snapshot chunk to a given follower
 729      *  InstallSnapshot should qualify as a heartbeat too.
 730      */
 731     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 732         try {
 733             if (snapshot.isPresent()) {
 734                 byte[] nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes());
 735
 736                 // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
 737                 // followerId to the followerToSnapshot map.
 738                 FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 739
 740                 int nextChunkIndex = followerToSnapshot.incrementChunkIndex();
 741                 Optional<ServerConfigurationPayload> serverConfig = Optional.absent();
 742                 if(followerToSnapshot.isLastChunk(nextChunkIndex)) {
 743                     serverConfig = Optional.fromNullable(context.getPeerServerInfo(true));
 744                 }
 745
 746                 followerActor.tell(
 747                     new InstallSnapshot(currentTerm(), context.getId(),
 748                         snapshot.get().getLastIncludedIndex(),
 749                         snapshot.get().getLastIncludedTerm(),
 750                         nextSnapshotChunk,
 751                         nextChunkIndex,
 752                         followerToSnapshot.getTotalChunks(),
 753                         Optional.of(followerToSnapshot.getLastChunkHashCode()),
 754                         serverConfig
 755                     ).toSerializable(followerToLog.get(followerId).getRaftVersion()),
 756                     actor()
 757                 );
 758
 759                 if(LOG.isDebugEnabled()) {
 760                     LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
 761                             logName(), followerActor.path(), followerToSnapshot.getChunkIndex(),
 762                             followerToSnapshot.getTotalChunks());
 763                 }
 764             }
 765         } catch (IOException e) {
 766             LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
 767         }
 768     }
 769
 770     /**
 771      * Acccepts snaphot as ByteString, enters into map for future chunks
 772      * creates and return a ByteString chunk
 773      */
 774     private byte[] getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 775         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 776         if (followerToSnapshot == null) {
 777             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 778             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 779         }
 780         byte[] nextChunk = followerToSnapshot.getNextChunk();
 781
 782         LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.length);
 783
 784         return nextChunk;
 785     }
 786
 787     private void sendHeartBeat() {
 788         if (!followerToLog.isEmpty()) {
 789             LOG.trace("{}: Sending heartbeat", logName());
 790             sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toMillis(), true);
 791         }
 792     }
 793
 794     private void stopHeartBeat() {
 795         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 796             heartbeatSchedule.cancel();
 797         }
 798     }
 799
 800     private void scheduleHeartBeat(FiniteDuration interval) {
 801         if (followerToLog.isEmpty()) {
 802             // Optimization - do not bother scheduling a heartbeat as there are
 803             // no followers
 804             return;
 805         }
 806
 807         stopHeartBeat();
 808
 809         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 810         // message is sent to itself.
 811         // Scheduling the heartbeat only once here because heartbeats do not
 812         // need to be sent if there are other messages being sent to the remote
 813         // actor.
 814         heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
 815             interval, context.getActor(), SendHeartBeat.INSTANCE,
 816             context.getActorSystem().dispatcher(), context.getActor());
 817     }
 818
 819     @Override
 820     public void close() {
 821         stopHeartBeat();
 822     }
 823
 824     @Override
 825     public final String getLeaderId() {
 826         return context.getId();
 827     }
 828
 829     @Override
 830     public final short getLeaderPayloadVersion() {
 831         return context.getPayloadVersion();
 832     }
 833
 834     protected boolean isLeaderIsolated() {
 835         int minPresent = getMinIsolatedLeaderPeerCount();
 836         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 837             final PeerInfo peerInfo = context.getPeerInfo(followerLogInformation.getId());
 838             if(peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) {
 839                 --minPresent;
 840                 if (minPresent == 0) {
 841                     return false;
 842                 }
 843             }
 844         }
 845         return minPresent != 0;
 846     }
 847
 848     /**
 849      * Encapsulates the snapshot bytestring and handles the logic of sending
 850      * snapshot chunks
 851      */
 852     protected class FollowerToSnapshot {
 853         private final ByteString snapshotBytes;
 854         private int offset = 0;
 855         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 856         private int replyReceivedForOffset;
 857         // if replyStatus is false, the previous chunk is attempted
 858         private boolean replyStatus = false;
 859         private int chunkIndex;
 860         private final int totalChunks;
 861         private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 862         private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 863
 864         public FollowerToSnapshot(ByteString snapshotBytes) {
 865             this.snapshotBytes = snapshotBytes;
 866             int size = snapshotBytes.size();
 867             totalChunks = (size / context.getConfigParams().getSnapshotChunkSize()) +
 868                 (size % context.getConfigParams().getSnapshotChunkSize() > 0 ? 1 : 0);
 869             if(LOG.isDebugEnabled()) {
 870                 LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}",
 871                         logName(), size, totalChunks);
 872             }
 873             replyReceivedForOffset = -1;
 874             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 875         }
 876
 877         public ByteString getSnapshotBytes() {
 878             return snapshotBytes;
 879         }
 880
 881         public int incrementOffset() {
 882             if(replyStatus) {
 883                 // if prev chunk failed, we would want to sent the same chunk again
 884                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 885             }
 886             return offset;
 887         }
 888
 889         public int incrementChunkIndex() {
 890             if (replyStatus) {
 891                 // if prev chunk failed, we would want to sent the same chunk again
 892                 chunkIndex =  chunkIndex + 1;
 893             }
 894             return chunkIndex;
 895         }
 896
 897         public int getChunkIndex() {
 898             return chunkIndex;
 899         }
 900
 901         public int getTotalChunks() {
 902             return totalChunks;
 903         }
 904
 905         public boolean canSendNextChunk() {
 906             // we only send a false if a chunk is sent but we have not received a reply yet
 907             return replyReceivedForOffset == offset;
 908         }
 909
 910         public boolean isLastChunk(int chunkIndex) {
 911             return totalChunks == chunkIndex;
 912         }
 913
 914         public void markSendStatus(boolean success) {
 915             if (success) {
 916                 // if the chunk sent was successful
 917                 replyReceivedForOffset = offset;
 918                 replyStatus = true;
 919                 lastChunkHashCode = nextChunkHashCode;
 920             } else {
 921                 // if the chunk sent was failure
 922                 replyReceivedForOffset = offset;
 923                 replyStatus = false;
 924             }
 925         }
 926
 927         public byte[] getNextChunk() {
 928             int snapshotLength = getSnapshotBytes().size();
 929             int start = incrementOffset();
 930             int size = context.getConfigParams().getSnapshotChunkSize();
 931             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 932                 size = snapshotLength;
 933             } else if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 934                 size = snapshotLength - start;
 935             }
 936
 937             byte[] nextChunk = new byte[size];
 938             getSnapshotBytes().copyTo(nextChunk, start, 0, size);
 939             nextChunkHashCode = Arrays.hashCode(nextChunk);
 940
 941             LOG.debug("{}: Next chunk: total length={}, offset={}, size={}, hashCode={}", logName(),
 942                     snapshotLength, start, size, nextChunkHashCode);
 943             return nextChunk;
 944         }
 945
 946         /**
 947          * reset should be called when the Follower needs to be sent the snapshot from the beginning
 948          */
 949         public void reset(){
 950             offset = 0;
 951             replyStatus = false;
 952             replyReceivedForOffset = offset;
 953             chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX;
 954             lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE;
 955         }
 956
 957         public int getLastChunkHashCode() {
 958             return lastChunkHashCode;
 959         }
 960     }
 961
 962     // called from example-actor for printing the follower-states
 963     public String printFollowerStates() {
 964         final StringBuilder sb = new StringBuilder();
 965
 966         sb.append('[');
 967         for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
 968             sb.append('{');
 969             sb.append(followerLogInformation.getId());
 970             sb.append(" state:");
 971             sb.append(followerLogInformation.isFollowerActive());
 972             sb.append("},");
 973         }
 974         sb.append(']');
 975
 976         return sb.toString();
 977     }
 978
 979     @VisibleForTesting
 980     public FollowerLogInformation getFollower(String followerId) {
 981         return followerToLog.get(followerId);
 982     }
 983
 984     @VisibleForTesting
 985     protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) {
 986         mapFollowerToSnapshot.put(followerId, snapshot);
 987     }
 988
 989     @VisibleForTesting
 990     public int followerSnapshotSize() {
 991         return mapFollowerToSnapshot.size();
 992     }
 993
 994     @VisibleForTesting
 995     public int followerLogSize() {
 996         return followerToLog.size();
 997     }
 998
 999     private static class SnapshotHolder {
1000         private final long lastIncludedTerm;
1001         private final long lastIncludedIndex;
1002         private final ByteString snapshotBytes;
1003
1004         SnapshotHolder(Snapshot snapshot) {
1005             this.lastIncludedTerm = snapshot.getLastAppliedTerm();
1006             this.lastIncludedIndex = snapshot.getLastAppliedIndex();
1007             this.snapshotBytes = ByteString.copyFrom(snapshot.getState());
1008         }
1009
1010         long getLastIncludedTerm() {
1011             return lastIncludedTerm;
1012         }
1013
1014         long getLastIncludedIndex() {
1015             return lastIncludedIndex;
1016         }
1017
1018         ByteString getSnapshotBytes() {
1019             return snapshotBytes;
1020         }
1021     }
1022 }