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.HashSet;
  22 import java.util.List;
  23 import java.util.Optional;
  24 import java.util.Set;
  25 import java.util.concurrent.TimeUnit;
  26 import java.util.concurrent.atomic.AtomicBoolean;
  27 import java.util.function.Consumer;
  28 import org.eclipse.jdt.annotation.Nullable;
  29 import org.opendaylight.controller.cluster.messaging.MessageAssembler;
  30 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  31 import org.opendaylight.controller.cluster.raft.RaftState;
  32 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  33 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
  34 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
  35 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
  36 import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow;
  37 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  38 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  39 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  40 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  41 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  42 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
  43 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  44 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
  45 import org.opendaylight.controller.cluster.raft.persisted.Snapshot;
  46
  47 /**
  48  * The behavior of a RaftActor in the Follower raft state.
  49  * <ul>
  50  * <li> Respond to RPCs from candidates and leaders
  51  * <li> If election timeout elapses without receiving AppendEntries
  52  * RPC from current leader or granting vote to candidate:
  53  * convert to candidate
  54  * </ul>
  55  */
  56 public class Follower extends AbstractRaftActorBehavior {
  57     private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
  58
  59     private final SyncStatusTracker initialSyncStatusTracker;
  60
  61     private final MessageAssembler appendEntriesMessageAssembler;
  62
  63     private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted();
  64     private SnapshotTracker snapshotTracker = null;
  65     private String leaderId;
  66     private short leaderPayloadVersion;
  67
  68     public Follower(final RaftActorContext context) {
  69         this(context, null, (short)-1);
  70     }
  71
  72     public Follower(final RaftActorContext context, final String initialLeaderId,
  73             final short initialLeaderPayloadVersion) {
  74         super(context, RaftState.Follower);
  75         this.leaderId = initialLeaderId;
  76         this.leaderPayloadVersion = initialLeaderPayloadVersion;
  77
  78         initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), context.getConfigParams()
  79             .getSyncIndexThreshold());
  80
  81         appendEntriesMessageAssembler = MessageAssembler.builder().logContext(logName())
  82                 .fileBackedStreamFactory(context.getFileBackedOutputStreamFactory())
  83                 .assembledMessageCallback((message, sender) -> handleMessage(sender, message)).build();
  84
  85         if (context.getPeerIds().isEmpty() && getLeaderId() == null) {
  86             actor().tell(TimeoutNow.INSTANCE, actor());
  87         } else {
  88             scheduleElection(electionDuration());
  89         }
  90     }
  91
  92     @Override
  93     public final String getLeaderId() {
  94         return leaderId;
  95     }
  96
  97     @VisibleForTesting
  98     protected final void setLeaderId(final @Nullable String leaderId) {
  99         this.leaderId = leaderId;
 100     }
 101
 102     @Override
 103     public short getLeaderPayloadVersion() {
 104         return leaderPayloadVersion;
 105     }
 106
 107     @VisibleForTesting
 108     protected final void setLeaderPayloadVersion(final short leaderPayloadVersion) {
 109         this.leaderPayloadVersion = leaderPayloadVersion;
 110     }
 111
 112     private void restartLastLeaderMessageTimer() {
 113         if (lastLeaderMessageTimer.isRunning()) {
 114             lastLeaderMessageTimer.reset();
 115         }
 116
 117         lastLeaderMessageTimer.start();
 118     }
 119
 120     private boolean isLogEntryPresent(final long index) {
 121         if (context.getReplicatedLog().isInSnapshot(index)) {
 122             return true;
 123         }
 124
 125         ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
 126         return entry != null;
 127
 128     }
 129
 130     private void updateInitialSyncStatus(final long currentLeaderCommit, final String newLeaderId) {
 131         initialSyncStatusTracker.update(newLeaderId, currentLeaderCommit, context.getCommitIndex());
 132     }
 133
 134     @Override
 135     protected RaftActorBehavior handleAppendEntries(final ActorRef sender, final AppendEntries appendEntries) {
 136         int numLogEntries = appendEntries.getEntries().size();
 137         if (log.isTraceEnabled()) {
 138             log.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
 139         } else if (log.isDebugEnabled() && numLogEntries > 0) {
 140             log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 141         }
 142
 143         if (snapshotTracker != null && !snapshotTracker.getLeaderId().equals(appendEntries.getLeaderId())) {
 144             log.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the "
 145                 + "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId());
 146             closeSnapshotTracker();
 147         }
 148
 149         if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
 150             // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
 151             AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
 152                     lastIndex(), lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
 153                     appendEntries.getLeaderRaftVersion());
 154
 155             log.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
 156             sender.tell(reply, actor());
 157
 158             return this;
 159         }
 160
 161         // If we got here then we do appear to be talking to the leader
 162         leaderId = appendEntries.getLeaderId();
 163         leaderPayloadVersion = appendEntries.getPayloadVersion();
 164
 165         if (appendEntries.getLeaderAddress().isPresent()) {
 166             final String address = appendEntries.getLeaderAddress().get();
 167             log.debug("New leader address: {}", address);
 168
 169             context.setPeerAddress(leaderId, address);
 170             context.getConfigParams().getPeerAddressResolver().setResolved(leaderId, address);
 171         }
 172
 173         // First check if the logs are in sync or not
 174         if (isOutOfSync(appendEntries, sender)) {
 175             updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 176             return this;
 177         }
 178
 179         if (!processNewEntries(appendEntries, sender)) {
 180             updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 181             return this;
 182         }
 183
 184         long lastIndex = lastIndex();
 185         long prevCommitIndex = context.getCommitIndex();
 186
 187         // If leaderCommit > commitIndex, set commitIndex = min(leaderCommit, index of last new entry)
 188         if (appendEntries.getLeaderCommit() > prevCommitIndex) {
 189             context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
 190         }
 191
 192         if (prevCommitIndex != context.getCommitIndex()) {
 193             log.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
 194         }
 195
 196         AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
 197                 lastIndex, lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
 198                 appendEntries.getLeaderRaftVersion());
 199
 200         if (log.isTraceEnabled()) {
 201             log.trace("{}: handleAppendEntries returning : {}", logName(), reply);
 202         } else if (log.isDebugEnabled() && numLogEntries > 0) {
 203             log.debug("{}: handleAppendEntries returning : {}", logName(), reply);
 204         }
 205
 206         // Reply to the leader before applying any previous state so as not to hold up leader consensus.
 207         sender.tell(reply, actor());
 208
 209         updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 210
 211         // If leaderCommit > lastApplied, increment lastApplied and apply log[lastApplied] to state machine (§5.3).
 212         // lastApplied can be equal to lastIndex.
 213         if (appendEntries.getLeaderCommit() > context.getLastApplied() && context.getLastApplied() < lastIndex) {
 214             if (log.isDebugEnabled()) {
 215                 log.debug("{}: applyLogToStateMachine, appendEntries.getLeaderCommit(): {}, "
 216                         + "context.getLastApplied(): {}, lastIndex(): {}", logName(),
 217                     appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
 218             }
 219
 220             applyLogToStateMachine(appendEntries.getLeaderCommit());
 221         }
 222
 223         if (!context.getSnapshotManager().isCapturing()) {
 224             super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
 225         }
 226
 227         appendEntriesMessageAssembler.checkExpiredAssembledMessageState();
 228
 229         return this;
 230     }
 231
 232     private boolean processNewEntries(final AppendEntries appendEntries, final ActorRef sender) {
 233         int numLogEntries = appendEntries.getEntries().size();
 234         if (numLogEntries == 0) {
 235             return true;
 236         }
 237
 238         log.debug("{}: Number of entries to be appended = {}", logName(), numLogEntries);
 239
 240         long lastIndex = lastIndex();
 241         int addEntriesFrom = 0;
 242
 243         // First check for conflicting entries. If an existing entry conflicts with a new one (same index but different
 244         // term), delete the existing entry and all that follow it (§5.3)
 245         if (context.getReplicatedLog().size() > 0) {
 246             // Find the entry up until the one that is not in the follower's log
 247             for (int i = 0;i < numLogEntries; i++, addEntriesFrom++) {
 248                 ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
 249
 250                 if (!isLogEntryPresent(matchEntry.getIndex())) {
 251                     // newEntry not found in the log
 252                     break;
 253                 }
 254
 255                 long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex());
 256
 257                 log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry,
 258                         existingEntryTerm);
 259
 260                 // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know
 261                 // what the term was so we'll assume it matches.
 262                 if (existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) {
 263                     continue;
 264                 }
 265
 266                 if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
 267                     log.info("{}: Removing entries from log starting at {}, commitIndex: {}, lastApplied: {}",
 268                             logName(), matchEntry.getIndex(), context.getCommitIndex(), context.getLastApplied());
 269
 270                     // Entries do not match so remove all subsequent entries but only if the existing entries haven't
 271                     // been applied to the state yet.
 272                     if (matchEntry.getIndex() <= context.getLastApplied()
 273                             || !context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) {
 274                         // Could not remove the entries - this means the matchEntry index must be in the
 275                         // snapshot and not the log. In this case the prior entries are part of the state
 276                         // so we must send back a reply to force a snapshot to completely re-sync the
 277                         // follower's log and state.
 278
 279                         log.info("{}: Could not remove entries - sending reply to force snapshot", logName());
 280                         sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
 281                                 lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
 282                                 appendEntries.getLeaderRaftVersion()), actor());
 283                         return false;
 284                     }
 285
 286                     break;
 287                 } else {
 288                     sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
 289                             lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
 290                             appendEntries.getLeaderRaftVersion()), actor());
 291                     return false;
 292                 }
 293             }
 294         }
 295
 296         lastIndex = lastIndex();
 297         log.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(), lastIndex,
 298                 addEntriesFrom);
 299
 300         // When persistence successfully completes for each new log entry appended, we need to determine if we
 301         // should capture a snapshot to compact the persisted log. shouldCaptureSnapshot tracks whether or not
 302         // one of the log entries has exceeded the log size threshold whereby a snapshot should be taken. However
 303         // we don't initiate the snapshot at that log entry but rather after the last log entry has been persisted.
 304         // This is done because subsequent log entries after the one that tripped the threshold may have been
 305         // applied to the state already, as the persistence callback occurs async, and we want those entries
 306         // purged from the persisted log as well.
 307         final AtomicBoolean shouldCaptureSnapshot = new AtomicBoolean(false);
 308         final Consumer<ReplicatedLogEntry> appendAndPersistCallback = logEntry -> {
 309             final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
 310             final ReplicatedLogEntry lastEntryToAppend = entries.get(entries.size() - 1);
 311             if (shouldCaptureSnapshot.get() && logEntry == lastEntryToAppend) {
 312                 context.getSnapshotManager().capture(context.getReplicatedLog().last(), getReplicatedToAllIndex());
 313             }
 314         };
 315
 316         // Append any new entries not already in the log
 317         for (int i = addEntriesFrom; i < numLogEntries; i++) {
 318             ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
 319
 320             log.debug("{}: Append entry to log {}", logName(), entry.getData());
 321
 322             context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback, false);
 323
 324             shouldCaptureSnapshot.compareAndSet(false,
 325                     context.getReplicatedLog().shouldCaptureSnapshot(entry.getIndex()));
 326
 327             if (entry.getData() instanceof ServerConfigurationPayload) {
 328                 context.updatePeerIds((ServerConfigurationPayload)entry.getData());
 329             }
 330         }
 331
 332         log.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
 333
 334         return true;
 335     }
 336
 337     private boolean isOutOfSync(final AppendEntries appendEntries, final ActorRef sender) {
 338
 339         final long lastIndex = lastIndex();
 340         if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
 341
 342             // The follower's log is out of sync because the leader does have an entry at prevLogIndex and this
 343             // follower has no entries in it's log.
 344
 345             log.info("{}: The followers log is empty and the senders prevLogIndex is {}", logName(),
 346                 appendEntries.getPrevLogIndex());
 347
 348             sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 349             return true;
 350         }
 351
 352         if (lastIndex > -1) {
 353             if (isLogEntryPresent(appendEntries.getPrevLogIndex())) {
 354                 final long leadersPrevLogTermInFollowersLogOrSnapshot =
 355                         getLogEntryOrSnapshotTerm(appendEntries.getPrevLogIndex());
 356                 if (leadersPrevLogTermInFollowersLogOrSnapshot != appendEntries.getPrevLogTerm()) {
 357
 358                     // The follower's log is out of sync because the Leader's prevLogIndex entry does exist
 359                     // in the follower's log or snapshot but it has a different term.
 360
 361                     log.info("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append "
 362                         + "entries prevLogTerm {} - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
 363                         appendEntries.getPrevLogIndex(), leadersPrevLogTermInFollowersLogOrSnapshot,
 364                         appendEntries.getPrevLogTerm(), lastIndex, context.getReplicatedLog().getSnapshotIndex(),
 365                         context.getReplicatedLog().getSnapshotTerm());
 366
 367                     sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 368                     return true;
 369                 }
 370             } else if (appendEntries.getPrevLogIndex() != -1) {
 371
 372                 // The follower's log is out of sync because the Leader's prevLogIndex entry was not found in it's log
 373
 374                 log.info("{}: The log is not empty but the prevLogIndex {} was not found in it - lastIndex: {}, "
 375                         + "snapshotIndex: {}, snapshotTerm: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex,
 376                         context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
 377
 378                 sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 379                 return true;
 380             }
 381         }
 382
 383         if (appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
 384                 && appendEntries.getReplicatedToAllIndex() != -1) {
 385             if (!isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
 386                 // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
 387                 // the previous entry in it's in-memory journal
 388
 389                 log.info("{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the "
 390                         + "in-memory journal - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
 391                         appendEntries.getReplicatedToAllIndex(), lastIndex,
 392                         context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
 393
 394                 sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 395                 return true;
 396             }
 397
 398             final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
 399             if (entries.size() > 0 && !isLogEntryPresent(entries.get(0).getIndex() - 1)) {
 400                 log.info("{}: Cannot append entries because the calculated previousIndex {} was not found in the "
 401                         + "in-memory journal - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
 402                         entries.get(0).getIndex() - 1, lastIndex, context.getReplicatedLog().getSnapshotIndex(),
 403                         context.getReplicatedLog().getSnapshotTerm());
 404
 405                 sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
 406                 return true;
 407             }
 408         }
 409
 410         return false;
 411     }
 412
 413     private void sendOutOfSyncAppendEntriesReply(final ActorRef sender, final boolean forceInstallSnapshot,
 414             final short leaderRaftVersion) {
 415         // We found that the log was out of sync so just send a negative reply.
 416         final AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex(),
 417                 lastTerm(), context.getPayloadVersion(), forceInstallSnapshot, needsLeaderAddress(),
 418                 leaderRaftVersion);
 419
 420         log.info("{}: Follower is out-of-sync so sending negative reply: {}", logName(), reply);
 421         sender.tell(reply, actor());
 422     }
 423
 424     private boolean needsLeaderAddress() {
 425         return context.getPeerAddress(leaderId) == null;
 426     }
 427
 428     @Override
 429     protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
 430         final AppendEntriesReply appendEntriesReply) {
 431         return this;
 432     }
 433
 434     @Override
 435     protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
 436         final RequestVoteReply requestVoteReply) {
 437         return this;
 438     }
 439
 440     @Override
 441     final ApplyState getApplyStateFor(final ReplicatedLogEntry entry) {
 442         return new ApplyState(null, null, entry);
 443     }
 444
 445     @Override
 446     public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
 447         if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
 448             return handleElectionTimeout(message);
 449         }
 450
 451         if (appendEntriesMessageAssembler.handleMessage(message, actor())) {
 452             return this;
 453         }
 454
 455         if (!(message instanceof RaftRPC)) {
 456             // The rest of the processing requires the message to be a RaftRPC
 457             return null;
 458         }
 459
 460         final RaftRPC rpc = (RaftRPC) message;
 461         // If RPC request or response contains term T > currentTerm:
 462         // set currentTerm = T, convert to follower (§5.1)
 463         // This applies to all RPC messages and responses
 464         if (rpc.getTerm() > context.getTermInformation().getCurrentTerm() && shouldUpdateTerm(rpc)) {
 465             log.info("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
 466                 logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 467
 468             context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 469         }
 470
 471         if (rpc instanceof InstallSnapshot) {
 472             handleInstallSnapshot(sender, (InstallSnapshot) rpc);
 473             restartLastLeaderMessageTimer();
 474             scheduleElection(electionDuration());
 475             return this;
 476         }
 477
 478         if (!(rpc instanceof RequestVote) || canGrantVote((RequestVote) rpc)) {
 479             restartLastLeaderMessageTimer();
 480             scheduleElection(electionDuration());
 481         }
 482
 483         return super.handleMessage(sender, rpc);
 484     }
 485
 486     private RaftActorBehavior handleElectionTimeout(final Object message) {
 487         // If the message is ElectionTimeout, verify we haven't actually seen a message from the leader
 488         // during the election timeout interval. It may that the election timer expired b/c this actor
 489         // was busy and messages got delayed, in which case leader messages would be backed up in the
 490         // queue but would be processed before the ElectionTimeout message and thus would restart the
 491         // lastLeaderMessageTimer.
 492         long lastLeaderMessageInterval = lastLeaderMessageTimer.elapsed(TimeUnit.MILLISECONDS);
 493         long electionTimeoutInMillis = context.getConfigParams().getElectionTimeOutInterval().toMillis();
 494         boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning()
 495                 || lastLeaderMessageInterval >= electionTimeoutInMillis;
 496
 497         if (canStartElection()) {
 498             if (message instanceof TimeoutNow) {
 499                 log.debug("{}: Received TimeoutNow - switching to Candidate", logName());
 500                 return internalSwitchBehavior(RaftState.Candidate);
 501             } else if (noLeaderMessageReceived) {
 502                 // Check the cluster state to see if the leader is known to be up before we go to Candidate.
 503                 // However if we haven't heard from the leader in a long time even though the cluster state
 504                 // indicates it's up then something is wrong - leader might be stuck indefinitely - so switch
 505                 // to Candidate,
 506                 long maxElectionTimeout = electionTimeoutInMillis * MAX_ELECTION_TIMEOUT_FACTOR;
 507                 if (isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
 508                     log.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
 509                     scheduleElection(electionDuration());
 510                 } else if (isThisFollowerIsolated()) {
 511                     log.debug("{}: this follower is isolated. Do not switch to Candidate for now.", logName());
 512                     setLeaderId(null);
 513                     scheduleElection(electionDuration());
 514                 } else {
 515                     log.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
 516                     return internalSwitchBehavior(RaftState.Candidate);
 517                 }
 518             } else {
 519                 log.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
 520                         logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
 521                 scheduleElection(electionDuration());
 522             }
 523         } else if (message instanceof ElectionTimeout) {
 524             if (noLeaderMessageReceived) {
 525                 setLeaderId(null);
 526             }
 527
 528             scheduleElection(electionDuration());
 529         }
 530
 531         return this;
 532     }
 533
 534     private boolean isLeaderAvailabilityKnown() {
 535         if (leaderId == null) {
 536             return false;
 537         }
 538
 539         Optional<Cluster> cluster = context.getCluster();
 540         if (!cluster.isPresent()) {
 541             return false;
 542         }
 543
 544         ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
 545         if (leaderActor == null) {
 546             return false;
 547         }
 548
 549         Address leaderAddress = leaderActor.anchorPath().address();
 550
 551         CurrentClusterState state = cluster.get().state();
 552         Set<Member> unreachable = state.getUnreachable();
 553
 554         log.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
 555                 unreachable);
 556
 557         for (Member m: unreachable) {
 558             if (leaderAddress.equals(m.address())) {
 559                 log.info("{}: Leader {} is unreachable", logName(), leaderAddress);
 560                 return false;
 561             }
 562         }
 563
 564         for (Member m: state.getMembers()) {
 565             if (leaderAddress.equals(m.address())) {
 566                 if (m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
 567                     log.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
 568                             leaderAddress, m.status());
 569                     return true;
 570                 } else {
 571                     log.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
 572                             leaderAddress, m.status());
 573                     return false;
 574                 }
 575             }
 576         }
 577
 578         log.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
 579
 580         return false;
 581     }
 582
 583     private boolean isThisFollowerIsolated() {
 584         final Optional<Cluster> maybeCluster = context.getCluster();
 585         if (!maybeCluster.isPresent()) {
 586             return false;
 587         }
 588
 589         final Cluster cluster = maybeCluster.get();
 590         final Member selfMember = cluster.selfMember();
 591
 592         final CurrentClusterState state = cluster.state();
 593         final Set<Member> unreachable = state.getUnreachable();
 594         final Iterable<Member> members = state.getMembers();
 595
 596         log.debug("{}: Checking if this node is isolated in the cluster unreachable set {},"
 597                         + "all members {} self member: {}", logName(), unreachable, members, selfMember);
 598
 599         // no unreachable peers means we cannot be isolated
 600         if (unreachable.size() == 0) {
 601             return false;
 602         }
 603
 604         final Set<Member> membersToCheck = new HashSet<>();
 605         members.forEach(membersToCheck::add);
 606
 607         membersToCheck.removeAll(unreachable);
 608
 609         // check if the only member not unreachable is us
 610         if (membersToCheck.size() == 1 && membersToCheck.iterator().next().equals(selfMember)) {
 611             return true;
 612         }
 613
 614         return false;
 615     }
 616
 617     private void handleInstallSnapshot(final ActorRef sender, final InstallSnapshot installSnapshot) {
 618
 619         log.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
 620
 621         leaderId = installSnapshot.getLeaderId();
 622
 623         if (snapshotTracker == null) {
 624             snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId(),
 625                     context);
 626         }
 627
 628         updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
 629
 630         try {
 631             final InstallSnapshotReply reply = new InstallSnapshotReply(
 632                     currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
 633
 634             if (snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
 635                     installSnapshot.getLastChunkHashCode())) {
 636
 637                 log.info("{}: Snapshot installed from leader: {}", logName(), installSnapshot.getLeaderId());
 638
 639                 Snapshot snapshot = Snapshot.create(
 640                         context.getSnapshotManager().convertSnapshot(snapshotTracker.getSnapshotBytes()),
 641                         new ArrayList<>(),
 642                         installSnapshot.getLastIncludedIndex(),
 643                         installSnapshot.getLastIncludedTerm(),
 644                         installSnapshot.getLastIncludedIndex(),
 645                         installSnapshot.getLastIncludedTerm(),
 646                         context.getTermInformation().getCurrentTerm(),
 647                         context.getTermInformation().getVotedFor(),
 648                         installSnapshot.getServerConfig().orElse(null));
 649
 650                 ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
 651                     @Override
 652                     public void onSuccess() {
 653                         log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 654
 655                         sender.tell(reply, actor());
 656                     }
 657
 658                     @Override
 659                     public void onFailure() {
 660                         sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor());
 661                     }
 662                 };
 663
 664                 actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
 665
 666                 closeSnapshotTracker();
 667             } else {
 668                 log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 669
 670                 sender.tell(reply, actor());
 671             }
 672         } catch (IOException e) {
 673             log.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
 674
 675             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
 676                     -1, false), actor());
 677
 678             closeSnapshotTracker();
 679         }
 680     }
 681
 682     private void closeSnapshotTracker() {
 683         if (snapshotTracker != null) {
 684             snapshotTracker.close();
 685             snapshotTracker = null;
 686         }
 687     }
 688
 689     @Override
 690     public void close() {
 691         closeSnapshotTracker();
 692         stopElection();
 693         appendEntriesMessageAssembler.close();
 694     }
 695
 696     @VisibleForTesting
 697     SnapshotTracker getSnapshotTracker() {
 698         return snapshotTracker;
 699     }
 700 }