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