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

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