opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/Follower.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 package org.opendaylight.controller.cluster.raft.behaviors;
   9
  10 import akka.actor.ActorRef;
  11 import akka.actor.ActorSelection;
  12 import akka.actor.Address;
  13 import akka.cluster.Cluster;
  14 import akka.cluster.ClusterEvent.CurrentClusterState;
  15 import akka.cluster.Member;
  16 import akka.cluster.MemberStatus;
  17 import com.google.common.annotations.VisibleForTesting;
  18 import com.google.common.base.Stopwatch;
  19 import java.io.IOException;
  20 import java.util.ArrayList;
  21 import java.util.List;
  22 import java.util.Optional;
  23 import java.util.Set;
  24 import java.util.concurrent.TimeUnit;
  25 import java.util.concurrent.atomic.AtomicBoolean;
  26 import java.util.function.Consumer;
  27 import org.eclipse.jdt.annotation.Nullable;
  28 import org.opendaylight.controller.cluster.messaging.MessageAssembler;
  29 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  30 import org.opendaylight.controller.cluster.raft.RaftState;
  31 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  32 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
  33 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  34 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
  35 import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow;
  36 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  37 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  38 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  39 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  40 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  41 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
  42 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  43 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
  44 import org.opendaylight.controller.cluster.raft.persisted.Snapshot;
  45
  46 /**
  47  * The behavior of a RaftActor in the Follower raft state.
  48  * <ul>
  49  * <li> Respond to RPCs from candidates and leaders
  50  * <li> If election timeout elapses without receiving AppendEntries
  51  * RPC from current leader or granting vote to candidate:
  52  * convert to candidate
  53  * </ul>
  54  */
  55 public class Follower extends AbstractRaftActorBehavior {
  56     private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
  57
  58     private final SyncStatusTracker initialSyncStatusTracker;
  59
  60     private final MessageAssembler appendEntriesMessageAssembler;
  61
  62     private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted();
  63     private SnapshotTracker snapshotTracker = null;
  64     private String leaderId;
  65     private short leaderPayloadVersion;
  66
  67     public Follower(final RaftActorContext context) {
  68         this(context, null, (short)-1);
  69     }
  70
  71     public Follower(final RaftActorContext context, final String initialLeaderId,
  72             final short initialLeaderPayloadVersion) {
  73         super(context, RaftState.Follower);
  74         this.leaderId = initialLeaderId;
  75         this.leaderPayloadVersion = initialLeaderPayloadVersion;
  76
  77         initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), context.getConfigParams()
  78             .getSyncIndexThreshold());
  79
  80         appendEntriesMessageAssembler = MessageAssembler.builder().logContext(logName())
  81                 .fileBackedStreamFactory(context.getFileBackedOutputStreamFactory())
  82                 .assembledMessageCallback((message, sender) -> handleMessage(sender, message)).build();
  83
  84         if (context.getPeerIds().isEmpty() && getLeaderId() == null) {
  85             actor().tell(TimeoutNow.INSTANCE, actor());
  86         } else {
  87             scheduleElection(electionDuration());
  88         }
  89     }
  90
  91     @Override
  92     public final String getLeaderId() {
  93         return leaderId;
  94     }
  95
  96     @VisibleForTesting
  97     protected final void setLeaderId(final @Nullable String leaderId) {
  98         this.leaderId = leaderId;
  99     }
 100
 101     @Override
 102     public short getLeaderPayloadVersion() {
 103         return leaderPayloadVersion;
 104     }
 105
 106     @VisibleForTesting
 107     protected final void setLeaderPayloadVersion(final short leaderPayloadVersion) {
 108         this.leaderPayloadVersion = leaderPayloadVersion;
 109     }
 110
 111     private void restartLastLeaderMessageTimer() {
 112         if (lastLeaderMessageTimer.isRunning()) {
 113             lastLeaderMessageTimer.reset();
 114         }
 115
 116         lastLeaderMessageTimer.start();
 117     }
 118
 119     private boolean isLogEntryPresent(final long index) {
 120         if (context.getReplicatedLog().isInSnapshot(index)) {
 121             return true;
 122         }
 123
 124         ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
 125         return entry != null;
 126
 127     }
 128
 129     private void updateInitialSyncStatus(final long currentLeaderCommit, final String newLeaderId) {
 130         initialSyncStatusTracker.update(newLeaderId, currentLeaderCommit, context.getCommitIndex());
 131     }
 132
 133     @Override
 134     protected RaftActorBehavior handleAppendEntries(final ActorRef sender, final AppendEntries appendEntries) {
 135         int numLogEntries = appendEntries.getEntries().size();
 136         if (log.isTraceEnabled()) {
 137             log.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
 138         } else if (log.isDebugEnabled() && numLogEntries > 0) {
 139             log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 140         }
 141
 142         if (snapshotTracker != null && !snapshotTracker.getLeaderId().equals(appendEntries.getLeaderId())) {
 143             log.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the "
 144                 + "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId());
 145             closeSnapshotTracker();
 146         }
 147
 148         if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
 149             // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
 150             AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
 151                     lastIndex(), lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
 152                     appendEntries.getLeaderRaftVersion());
 153
 154             log.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
 155             sender.tell(reply, actor());
 156
 157             return this;
 158         }
 159
 160         // If we got here then we do appear to be talking to the leader
 161         leaderId = appendEntries.getLeaderId();
 162         leaderPayloadVersion = appendEntries.getPayloadVersion();
 163
 164         if (appendEntries.getLeaderAddress().isPresent()) {
 165             final String address = appendEntries.getLeaderAddress().get();
 166             log.debug("New leader address: {}", address);
 167
 168             context.setPeerAddress(leaderId, address);
 169             context.getConfigParams().getPeerAddressResolver().setResolved(leaderId, address);
 170         }
 171
 172         // First check if the logs are in sync or not
 173         if (isOutOfSync(appendEntries, sender)) {
 174             updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 175             return this;
 176         }
 177
 178         if (!processNewEntries(appendEntries, sender)) {
 179             updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 180             return this;
 181         }
 182
 183         long lastIndex = lastIndex();
 184         long prevCommitIndex = context.getCommitIndex();
 185
 186         // If leaderCommit > commitIndex, set commitIndex = min(leaderCommit, index of last new entry)
 187         if (appendEntries.getLeaderCommit() > prevCommitIndex) {
 188             context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
 189         }
 190
 191         if (prevCommitIndex != context.getCommitIndex()) {
 192             log.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
 193         }
 194
 195         AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
 196                 lastIndex, lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
 197                 appendEntries.getLeaderRaftVersion());
 198
 199         if (log.isTraceEnabled()) {
 200             log.trace("{}: handleAppendEntries returning : {}", logName(), reply);
 201         } else if (log.isDebugEnabled() && numLogEntries > 0) {
 202             log.debug("{}: handleAppendEntries returning : {}", logName(), reply);
 203         }
 204
 205         // Reply to the leader before applying any previous state so as not to hold up leader consensus.
 206         sender.tell(reply, actor());
 207
 208         updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 209
 210         // If leaderCommit > lastApplied, increment lastApplied and apply log[lastApplied] to state machine (§5.3).
 211         // lastApplied can be equal to lastIndex.
 212         if (appendEntries.getLeaderCommit() > context.getLastApplied() && context.getLastApplied() < lastIndex) {
 213             if (log.isDebugEnabled()) {
 214                 log.debug("{}: applyLogToStateMachine, appendEntries.getLeaderCommit(): {}, "
 215                         + "context.getLastApplied(): {}, lastIndex(): {}", logName(),
 216                     appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
 217             }
 218
 219             applyLogToStateMachine(appendEntries.getLeaderCommit());
 220         }
 221
 222         if (!context.getSnapshotManager().isCapturing()) {
 223             super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
 224         }
 225
 226         appendEntriesMessageAssembler.checkExpiredAssembledMessageState();
 227
 228         return this;
 229     }
 230
 231     private boolean processNewEntries(final AppendEntries appendEntries, final ActorRef sender) {
 232         int numLogEntries = appendEntries.getEntries().size();
 233         if (numLogEntries == 0) {
 234             return true;
 235         }
 236
 237         log.debug("{}: Number of entries to be appended = {}", logName(), numLogEntries);
 238
 239         long lastIndex = lastIndex();
 240         int addEntriesFrom = 0;
 241
 242         // First check for conflicting entries. If an existing entry conflicts with a new one (same index but different
 243         // term), delete the existing entry and all that follow it (§5.3)
 244         if (context.getReplicatedLog().size() > 0) {
 245             // Find the entry up until the one that is not in the follower's log
 246             for (int i = 0;i < numLogEntries; i++, addEntriesFrom++) {
 247                 ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
 248
 249                 if (!isLogEntryPresent(matchEntry.getIndex())) {
 250                     // newEntry not found in the log
 251                     break;
 252                 }
 253
 254                 long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex());
 255
 256                 log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry,
 257                         existingEntryTerm);
 258
 259                 // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know
 260                 // what the term was so we'll assume it matches.
 261                 if (existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) {
 262                     continue;
 263                 }
 264
 265                 if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
 266                     log.info("{}: Removing entries from log starting at {}, commitIndex: {}, lastApplied: {}",
 267                             logName(), matchEntry.getIndex(), context.getCommitIndex(), context.getLastApplied());
 268
 269                     // Entries do not match so remove all subsequent entries but only if the existing entries haven't
 270                     // been applied to the state yet.
 271                     if (matchEntry.getIndex() <= context.getLastApplied()
 272                             || !context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) {
 273                         // Could not remove the entries - this means the matchEntry index must be in the
 274                         // snapshot and not the log. In this case the prior entries are part of the state
 275                         // so we must send back a reply to force a snapshot to completely re-sync the
 276                         // follower's log and state.
 277
 278                         log.info("{}: Could not remove entries - sending reply to force snapshot", logName());
 279                         sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
 280                                 lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
 281                                 appendEntries.getLeaderRaftVersion()), actor());
 282                         return false;
 283                     }
 284
 285                     break;
 286                 } else {
 287                     sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
 288                             lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
 289                             appendEntries.getLeaderRaftVersion()), actor());
 290                     return false;
 291                 }
 292             }
 293         }
 294
 295         lastIndex = lastIndex();
 296         log.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(), lastIndex,
 297                 addEntriesFrom);
 298
 299         // When persistence successfully completes for each new log entry appended, we need to determine if we
 300         // should capture a snapshot to compact the persisted log. shouldCaptureSnapshot tracks whether or not
 301         // one of the log entries has exceeded the log size threshold whereby a snapshot should be taken. However
 302         // we don't initiate the snapshot at that log entry but rather after the last log entry has been persisted.
 303         // This is done because subsequent log entries after the one that tripped the threshold may have been
 304         // applied to the state already, as the persistence callback occurs async, and we want those entries
 305         // purged from the persisted log as well.
 306         final AtomicBoolean shouldCaptureSnapshot = new AtomicBoolean(false);
 307         final Consumer<ReplicatedLogEntry> appendAndPersistCallback = logEntry -> {
 308             final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
 309             final ReplicatedLogEntry lastEntryToAppend = entries.get(entries.size() - 1);
 310             if (shouldCaptureSnapshot.get() && logEntry == lastEntryToAppend) {
 311                 context.getSnapshotManager().capture(context.getReplicatedLog().last(), getReplicatedToAllIndex());
 312             }
 313         };
 314
 315         // Append any new entries not already in the log
 316         for (int i = addEntriesFrom; i < numLogEntries; i++) {
 317             ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
 318
 319             log.debug("{}: Append entry to log {}", logName(), entry.getData());
 320
 321             context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback, false);
 322
 323             shouldCaptureSnapshot.compareAndSet(false,
 324                     context.getReplicatedLog().shouldCaptureSnapshot(entry.getIndex()));
 325
 326             if (entry.getData() instanceof ServerConfigurationPayload) {
 327                 context.updatePeerIds((ServerConfigurationPayload)entry.getData());
 328             }
 329         }
 330
 331         log.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
 332
 333         return true;
 334     }
 335
 336     private boolean isOutOfSync(final AppendEntries appendEntries, final ActorRef sender) {
 337
 338         final long lastIndex = lastIndex();
 339         if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
 340
 341             // The follower's log is out of sync because the leader does have an entry at prevLogIndex and this
 342             // follower has no entries in it's log.
 343
 344             log.info("{}: The followers log is empty and the senders prevLogIndex is {}", logName(),
 345                 appendEntries.getPrevLogIndex());
 346
 347             sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 348             return true;
 349         }
 350
 351         if (lastIndex > -1) {
 352             if (isLogEntryPresent(appendEntries.getPrevLogIndex())) {
 353                 final long leadersPrevLogTermInFollowersLogOrSnapshot =
 354                         getLogEntryOrSnapshotTerm(appendEntries.getPrevLogIndex());
 355                 if (leadersPrevLogTermInFollowersLogOrSnapshot != appendEntries.getPrevLogTerm()) {
 356
 357                     // The follower's log is out of sync because the Leader's prevLogIndex entry does exist
 358                     // in the follower's log or snapshot but it has a different term.
 359
 360                     log.info("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append "
 361                         + "entries prevLogTerm {} - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
 362                         appendEntries.getPrevLogIndex(), leadersPrevLogTermInFollowersLogOrSnapshot,
 363                         appendEntries.getPrevLogTerm(), lastIndex, context.getReplicatedLog().getSnapshotIndex(),
 364                         context.getReplicatedLog().getSnapshotTerm());
 365
 366                     sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 367                     return true;
 368                 }
 369             } else if (appendEntries.getPrevLogIndex() != -1) {
 370
 371                 // The follower's log is out of sync because the Leader's prevLogIndex entry was not found in it's log
 372
 373                 log.info("{}: The log is not empty but the prevLogIndex {} was not found in it - lastIndex: {}, "
 374                         + "snapshotIndex: {}, snapshotTerm: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex,
 375                         context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
 376
 377                 sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 378                 return true;
 379             }
 380         }
 381
 382         if (appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
 383                 && appendEntries.getReplicatedToAllIndex() != -1) {
 384             if (!isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
 385                 // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
 386                 // the previous entry in it's in-memory journal
 387
 388                 log.info("{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the "
 389                         + "in-memory journal - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
 390                         appendEntries.getReplicatedToAllIndex(), lastIndex,
 391                         context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
 392
 393                 sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 394                 return true;
 395             }
 396
 397             final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
 398             if (entries.size() > 0 && !isLogEntryPresent(entries.get(0).getIndex() - 1)) {
 399                 log.info("{}: Cannot append entries because the calculated previousIndex {} was not found in the "
 400                         + "in-memory journal - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
 401                         entries.get(0).getIndex() - 1, lastIndex, context.getReplicatedLog().getSnapshotIndex(),
 402                         context.getReplicatedLog().getSnapshotTerm());
 403
 404                 sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 405                 return true;
 406             }
 407         }
 408
 409         return false;
 410     }
 411
 412     private void sendOutOfSyncAppendEntriesReply(final ActorRef sender, final boolean forceInstallSnapshot,
 413             final short leaderRaftVersion) {
 414         // We found that the log was out of sync so just send a negative reply.
 415         final AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex(),
 416                 lastTerm(), context.getPayloadVersion(), forceInstallSnapshot, needsLeaderAddress(),
 417                 leaderRaftVersion);
 418
 419         log.info("{}: Follower is out-of-sync so sending negative reply: {}", logName(), reply);
 420         sender.tell(reply, actor());
 421     }
 422
 423     private boolean needsLeaderAddress() {
 424         return context.getPeerAddress(leaderId) == null;
 425     }
 426
 427     @Override
 428     protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
 429         final AppendEntriesReply appendEntriesReply) {
 430         return this;
 431     }
 432
 433     @Override
 434     protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
 435         final RequestVoteReply requestVoteReply) {
 436         return this;
 437     }
 438
 439     @Override
 440     final ApplyState getApplyStateFor(final ReplicatedLogEntry entry) {
 441         return new ApplyState(null, null, entry);
 442     }
 443
 444     @Override
 445     public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
 446         if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
 447             return handleElectionTimeout(message);
 448         }
 449
 450         if (appendEntriesMessageAssembler.handleMessage(message, actor())) {
 451             return this;
 452         }
 453
 454         if (!(message instanceof RaftRPC)) {
 455             // The rest of the processing requires the message to be a RaftRPC
 456             return null;
 457         }
 458
 459         final RaftRPC rpc = (RaftRPC) message;
 460         // If RPC request or response contains term T > currentTerm:
 461         // set currentTerm = T, convert to follower (§5.1)
 462         // This applies to all RPC messages and responses
 463         if (rpc.getTerm() > context.getTermInformation().getCurrentTerm() && shouldUpdateTerm(rpc)) {
 464             log.info("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
 465                 logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 466
 467             context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 468         }
 469
 470         if (rpc instanceof InstallSnapshot) {
 471             handleInstallSnapshot(sender, (InstallSnapshot) rpc);
 472             restartLastLeaderMessageTimer();
 473             scheduleElection(electionDuration());
 474             return this;
 475         }
 476
 477         if (!(rpc instanceof RequestVote) || canGrantVote((RequestVote) rpc)) {
 478             restartLastLeaderMessageTimer();
 479             scheduleElection(electionDuration());
 480         }
 481
 482         return super.handleMessage(sender, rpc);
 483     }
 484
 485     private RaftActorBehavior handleElectionTimeout(final Object message) {
 486         // If the message is ElectionTimeout, verify we haven't actually seen a message from the leader
 487         // during the election timeout interval. It may that the election timer expired b/c this actor
 488         // was busy and messages got delayed, in which case leader messages would be backed up in the
 489         // queue but would be processed before the ElectionTimeout message and thus would restart the
 490         // lastLeaderMessageTimer.
 491         long lastLeaderMessageInterval = lastLeaderMessageTimer.elapsed(TimeUnit.MILLISECONDS);
 492         long electionTimeoutInMillis = context.getConfigParams().getElectionTimeOutInterval().toMillis();
 493         boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning()
 494                 || lastLeaderMessageInterval >= electionTimeoutInMillis;
 495
 496         if (canStartElection()) {
 497             if (message instanceof TimeoutNow) {
 498                 log.debug("{}: Received TimeoutNow - switching to Candidate", logName());
 499                 return internalSwitchBehavior(RaftState.Candidate);
 500             } else if (noLeaderMessageReceived) {
 501                 // Check the cluster state to see if the leader is known to be up before we go to Candidate.
 502                 // However if we haven't heard from the leader in a long time even though the cluster state
 503                 // indicates it's up then something is wrong - leader might be stuck indefinitely - so switch
 504                 // to Candidate,
 505                 long maxElectionTimeout = electionTimeoutInMillis * MAX_ELECTION_TIMEOUT_FACTOR;
 506                 if (isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
 507                     log.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
 508                     scheduleElection(electionDuration());
 509                 } else {
 510                     log.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
 511                     return internalSwitchBehavior(RaftState.Candidate);
 512                 }
 513             } else {
 514                 log.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
 515                         logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
 516                 scheduleElection(electionDuration());
 517             }
 518         } else if (message instanceof ElectionTimeout) {
 519             if (noLeaderMessageReceived) {
 520                 setLeaderId(null);
 521             }
 522
 523             scheduleElection(electionDuration());
 524         }
 525
 526         return this;
 527     }
 528
 529     private boolean isLeaderAvailabilityKnown() {
 530         if (leaderId == null) {
 531             return false;
 532         }
 533
 534         Optional<Cluster> cluster = context.getCluster();
 535         if (!cluster.isPresent()) {
 536             return false;
 537         }
 538
 539         ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
 540         if (leaderActor == null) {
 541             return false;
 542         }
 543
 544         Address leaderAddress = leaderActor.anchorPath().address();
 545
 546         CurrentClusterState state = cluster.get().state();
 547         Set<Member> unreachable = state.getUnreachable();
 548
 549         log.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
 550                 unreachable);
 551
 552         for (Member m: unreachable) {
 553             if (leaderAddress.equals(m.address())) {
 554                 log.info("{}: Leader {} is unreachable", logName(), leaderAddress);
 555                 return false;
 556             }
 557         }
 558
 559         for (Member m: state.getMembers()) {
 560             if (leaderAddress.equals(m.address())) {
 561                 if (m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
 562                     log.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
 563                             leaderAddress, m.status());
 564                     return true;
 565                 } else {
 566                     log.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
 567                             leaderAddress, m.status());
 568                     return false;
 569                 }
 570             }
 571         }
 572
 573         log.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
 574
 575         return false;
 576     }
 577
 578     private void handleInstallSnapshot(final ActorRef sender, final InstallSnapshot installSnapshot) {
 579
 580         log.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
 581
 582         leaderId = installSnapshot.getLeaderId();
 583
 584         if (snapshotTracker == null) {
 585             snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId(),
 586                     context);
 587         }
 588
 589         updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
 590
 591         try {
 592             final InstallSnapshotReply reply = new InstallSnapshotReply(
 593                     currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
 594
 595             if (snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
 596                     installSnapshot.getLastChunkHashCode())) {
 597
 598                 log.info("{}: Snapshot installed from leader: {}", logName(), installSnapshot.getLeaderId());
 599
 600                 Snapshot snapshot = Snapshot.create(
 601                         context.getSnapshotManager().convertSnapshot(snapshotTracker.getSnapshotBytes()),
 602                         new ArrayList<>(),
 603                         installSnapshot.getLastIncludedIndex(),
 604                         installSnapshot.getLastIncludedTerm(),
 605                         installSnapshot.getLastIncludedIndex(),
 606                         installSnapshot.getLastIncludedTerm(),
 607                         context.getTermInformation().getCurrentTerm(),
 608                         context.getTermInformation().getVotedFor(),
 609                         installSnapshot.getServerConfig().orElse(null));
 610
 611                 ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
 612                     @Override
 613                     public void onSuccess() {
 614                         log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 615
 616                         sender.tell(reply, actor());
 617                     }
 618
 619                     @Override
 620                     public void onFailure() {
 621                         sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor());
 622                     }
 623                 };
 624
 625                 actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
 626
 627                 closeSnapshotTracker();
 628             } else {
 629                 log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 630
 631                 sender.tell(reply, actor());
 632             }
 633         } catch (IOException e) {
 634             log.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
 635
 636             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
 637                     -1, false), actor());
 638
 639             closeSnapshotTracker();
 640         }
 641     }
 642
 643     private void closeSnapshotTracker() {
 644         if (snapshotTracker != null) {
 645             snapshotTracker.close();
 646             snapshotTracker = null;
 647         }
 648     }
 649
 650     @Override
 651     public void close() {
 652         closeSnapshotTracker();
 653         stopElection();
 654         appendEntriesMessageAssembler.close();
 655     }
 656
 657     @VisibleForTesting
 658     SnapshotTracker getSnapshotTracker() {
 659         return snapshotTracker;
 660     }
 661 }