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
   9 package org.opendaylight.controller.cluster.raft.behaviors;
  10
  11 import akka.actor.ActorRef;
  12 import com.google.common.annotations.VisibleForTesting;
  13 import java.util.ArrayList;
  14 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  15 import org.opendaylight.controller.cluster.raft.RaftState;
  16 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  17 import org.opendaylight.controller.cluster.raft.Snapshot;
  18 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
  19 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
  20 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  21 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  22 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  23 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  24 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  25 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  26 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
  27 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  28
  29 /**
  30  * The behavior of a RaftActor in the Follower state
  31  * <p/>
  32  * <ul>
  33  * <li> Respond to RPCs from candidates and leaders
  34  * <li> If election timeout elapses without receiving AppendEntries
  35  * RPC from current leader or granting vote to candidate:
  36  * convert to candidate
  37  * </ul>
  38  */
  39 public class Follower extends AbstractRaftActorBehavior {
  40
  41     private SnapshotTracker snapshotTracker = null;
  42
  43     private final InitialSyncStatusTracker initialSyncStatusTracker;
  44
  45     public Follower(RaftActorContext context) {
  46         super(context, RaftState.Follower);
  47
  48         initialSyncStatusTracker = new InitialSyncStatusTracker(context.getActor());
  49
  50         if(context.getRaftPolicy().automaticElectionsEnabled()) {
  51             if (context.getPeerAddresses().isEmpty()) {
  52                 actor().tell(ELECTION_TIMEOUT, actor());
  53             } else {
  54                 scheduleElection(electionDuration());
  55             }
  56         }
  57
  58     }
  59
  60     private boolean isLogEntryPresent(long index){
  61         if(index == context.getReplicatedLog().getSnapshotIndex()){
  62             return true;
  63         }
  64
  65         ReplicatedLogEntry previousEntry = context.getReplicatedLog()
  66                 .get(index);
  67
  68         return previousEntry != null;
  69
  70     }
  71
  72     private long getLogEntryTerm(long index){
  73         if(index == context.getReplicatedLog().getSnapshotIndex()){
  74             return context.getReplicatedLog().getSnapshotTerm();
  75         }
  76
  77         ReplicatedLogEntry previousEntry = context.getReplicatedLog()
  78                 .get(index);
  79
  80         if(previousEntry != null){
  81             return previousEntry.getTerm();
  82         }
  83
  84         return -1;
  85     }
  86
  87     private void updateInitialSyncStatus(long currentLeaderCommit, String leaderId){
  88         initialSyncStatusTracker.update(leaderId, currentLeaderCommit, context.getCommitIndex());
  89     }
  90
  91     @Override protected RaftActorBehavior handleAppendEntries(ActorRef sender,
  92                                                               AppendEntries appendEntries) {
  93
  94         int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
  95         if(LOG.isTraceEnabled()) {
  96             LOG.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
  97         } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
  98             LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
  99         }
 100
 101         // TODO : Refactor this method into a bunch of smaller methods
 102         // to make it easier to read. Before refactoring ensure tests
 103         // cover the code properly
 104
 105         if (snapshotTracker != null) {
 106             // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
 107             AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
 108                     lastIndex(), lastTerm(), context.getPayloadVersion());
 109
 110             if(LOG.isDebugEnabled()) {
 111                 LOG.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
 112             }
 113             sender.tell(reply, actor());
 114
 115             return this;
 116         }
 117
 118         // If we got here then we do appear to be talking to the leader
 119         leaderId = appendEntries.getLeaderId();
 120
 121         setLeaderPayloadVersion(appendEntries.getPayloadVersion());
 122
 123         updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
 124         // First check if the logs are in sync or not
 125         long lastIndex = lastIndex();
 126
 127         if (isOutOfSync(appendEntries)) {
 128             // We found that the log was out of sync so just send a negative
 129             // reply and return
 130
 131             LOG.debug("{}: Follower is out-of-sync, so sending negative reply, lastIndex: {}, lastTerm: {}",
 132                         logName(), lastIndex, lastTerm());
 133
 134             sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
 135                     lastTerm(), context.getPayloadVersion()), actor());
 136             return this;
 137         }
 138
 139         if (appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
 140
 141             LOG.debug("{}: Number of entries to be appended = {}", logName(),
 142                         appendEntries.getEntries().size());
 143
 144             // 3. If an existing entry conflicts with a new one (same index
 145             // but different terms), delete the existing entry and all that
 146             // follow it (§5.3)
 147             int addEntriesFrom = 0;
 148             if (context.getReplicatedLog().size() > 0) {
 149
 150                 // Find the entry up until which the one that is not in the follower's log
 151                 for (int i = 0;i < appendEntries.getEntries().size(); i++, addEntriesFrom++) {
 152                     ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
 153                     ReplicatedLogEntry newEntry = context.getReplicatedLog().get(matchEntry.getIndex());
 154
 155                     if (newEntry == null) {
 156                         //newEntry not found in the log
 157                         break;
 158                     }
 159
 160                     if (newEntry.getTerm() == matchEntry.getTerm()) {
 161                         continue;
 162                     }
 163
 164                     LOG.debug("{}: Removing entries from log starting at {}", logName(),
 165                                 matchEntry.getIndex());
 166
 167                     // Entries do not match so remove all subsequent entries
 168                     context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex());
 169                     break;
 170                 }
 171             }
 172
 173             lastIndex = lastIndex();
 174             LOG.debug("{}: After cleanup entries to be added from = {}", logName(),
 175                         (addEntriesFrom + lastIndex));
 176
 177             // 4. Append any new entries not already in the log
 178             for (int i = addEntriesFrom; i < appendEntries.getEntries().size(); i++) {
 179                 ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
 180
 181                 LOG.debug("{}: Append entry to log {}", logName(), entry.getData());
 182
 183                 context.getReplicatedLog().appendAndPersist(entry);
 184             }
 185
 186             LOG.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
 187         }
 188
 189         // 5. If leaderCommit > commitIndex, set commitIndex =
 190         // min(leaderCommit, index of last new entry)
 191
 192         lastIndex = lastIndex();
 193         long prevCommitIndex = context.getCommitIndex();
 194
 195         context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
 196
 197         if (prevCommitIndex != context.getCommitIndex()) {
 198             LOG.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
 199         }
 200
 201         // If commitIndex > lastApplied: increment lastApplied, apply
 202         // log[lastApplied] to state machine (§5.3)
 203         // check if there are any entries to be applied. last-applied can be equal to last-index
 204         if (appendEntries.getLeaderCommit() > context.getLastApplied() &&
 205             context.getLastApplied() < lastIndex) {
 206             if(LOG.isDebugEnabled()) {
 207                 LOG.debug("{}: applyLogToStateMachine, " +
 208                         "appendEntries.getLeaderCommit(): {}," +
 209                         "context.getLastApplied(): {}, lastIndex(): {}", logName(),
 210                     appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
 211             }
 212
 213             applyLogToStateMachine(appendEntries.getLeaderCommit());
 214         }
 215
 216         AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
 217             lastIndex, lastTerm(), context.getPayloadVersion());
 218
 219         if(LOG.isTraceEnabled()) {
 220             LOG.trace("{}: handleAppendEntries returning : {}", logName(), reply);
 221         } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
 222             LOG.debug("{}: handleAppendEntries returning : {}", logName(), reply);
 223         }
 224
 225         sender.tell(reply, actor());
 226
 227         if (!context.getSnapshotManager().isCapturing()) {
 228             super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
 229         }
 230
 231         return this;
 232     }
 233
 234     private boolean isOutOfSync(AppendEntries appendEntries) {
 235
 236         long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex());
 237         boolean prevEntryPresent = isLogEntryPresent(appendEntries.getPrevLogIndex());
 238         long lastIndex = lastIndex();
 239         int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
 240         boolean outOfSync = true;
 241
 242         if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
 243
 244             // The follower's log is out of sync because the leader does have
 245             // an entry at prevLogIndex and this follower has no entries in
 246             // it's log.
 247
 248             LOG.debug("{}: The followers log is empty and the senders prevLogIndex is {}",
 249                         logName(), appendEntries.getPrevLogIndex());
 250         } else if (lastIndex > -1 && appendEntries.getPrevLogIndex() != -1 && !prevEntryPresent) {
 251
 252             // The follower's log is out of sync because the Leader's
 253             // prevLogIndex entry was not found in it's log
 254
 255             LOG.debug("{}: The log is not empty but the prevLogIndex {} was not found in it",
 256                         logName(), appendEntries.getPrevLogIndex());
 257         } else if (lastIndex > -1 && prevEntryPresent && prevLogTerm != appendEntries.getPrevLogTerm()) {
 258
 259             // The follower's log is out of sync because the Leader's
 260             // prevLogIndex entry does exist in the follower's log but it has
 261             // a different term in it
 262
 263             LOG.debug(
 264                     "{}: Cannot append entries because previous entry term {}  is not equal to append entries prevLogTerm {}",
 265                     logName(), prevLogTerm, appendEntries.getPrevLogTerm());
 266         } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
 267                 && appendEntries.getReplicatedToAllIndex() != -1
 268                 && !isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
 269             // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
 270             // the previous entry in it's in-memory journal
 271
 272             LOG.debug(
 273                     "{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the in-memory journal",
 274                     logName(), appendEntries.getReplicatedToAllIndex());
 275         } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
 276                 && appendEntries.getReplicatedToAllIndex() != -1 && numLogEntries > 0 &&
 277                 !isLogEntryPresent(appendEntries.getEntries().get(0).getIndex() - 1)){
 278             LOG.debug(
 279                     "{}: Cannot append entries because the calculated previousIndex {} was not found in the in-memory journal",
 280                     logName(), appendEntries.getEntries().get(0).getIndex() - 1);
 281         } else {
 282             outOfSync = false;
 283         }
 284         return outOfSync;
 285     }
 286
 287     @Override protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 288         AppendEntriesReply appendEntriesReply) {
 289         return this;
 290     }
 291
 292     @Override protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 293         RequestVoteReply requestVoteReply) {
 294         return this;
 295     }
 296
 297     @Override public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 298
 299         Object message = fromSerializableMessage(originalMessage);
 300
 301         if (message instanceof RaftRPC) {
 302             RaftRPC rpc = (RaftRPC) message;
 303             // If RPC request or response contains term T > currentTerm:
 304             // set currentTerm = T, convert to follower (§5.1)
 305             // This applies to all RPC messages and responses
 306             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 307                 LOG.debug("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
 308                         logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 309
 310                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 311             }
 312         }
 313
 314         if (message instanceof ElectionTimeout) {
 315             LOG.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
 316             return switchBehavior(new Candidate(context));
 317
 318         } else if (message instanceof InstallSnapshot) {
 319             InstallSnapshot installSnapshot = (InstallSnapshot) message;
 320             handleInstallSnapshot(sender, installSnapshot);
 321         }
 322
 323         if(message instanceof RaftRPC && (!(message instanceof RequestVote) || (canGrantVote((RequestVote) message)))){
 324             scheduleElection(electionDuration());
 325         }
 326
 327         return super.handleMessage(sender, message);
 328     }
 329
 330     private void handleInstallSnapshot(ActorRef sender, InstallSnapshot installSnapshot) {
 331
 332         LOG.debug("{}: InstallSnapshot received from leader {}, datasize: {} , Chunk: {}/{}",
 333                     logName(), installSnapshot.getLeaderId(), installSnapshot.getData().size(),
 334                     installSnapshot.getChunkIndex(), installSnapshot.getTotalChunks());
 335
 336         if(snapshotTracker == null){
 337             snapshotTracker = new SnapshotTracker(LOG, installSnapshot.getTotalChunks());
 338         }
 339
 340         updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
 341
 342         try {
 343             if(snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
 344                     installSnapshot.getLastChunkHashCode())){
 345                 Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(),
 346                         new ArrayList<ReplicatedLogEntry>(),
 347                         installSnapshot.getLastIncludedIndex(),
 348                         installSnapshot.getLastIncludedTerm(),
 349                         installSnapshot.getLastIncludedIndex(),
 350                         installSnapshot.getLastIncludedTerm());
 351
 352                 actor().tell(new ApplySnapshot(snapshot), actor());
 353
 354                 snapshotTracker = null;
 355
 356             }
 357
 358             InstallSnapshotReply reply = new InstallSnapshotReply(
 359                     currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
 360
 361             LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 362
 363             sender.tell(reply, actor());
 364
 365         } catch (SnapshotTracker.InvalidChunkException e) {
 366             LOG.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
 367
 368             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
 369                     -1, false), actor());
 370             snapshotTracker = null;
 371
 372         } catch (Exception e){
 373             LOG.error("{}: Exception in InstallSnapshot of follower", logName(), e);
 374
 375             //send reply with success as false. The chunk will be sent again on failure
 376             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
 377                     installSnapshot.getChunkIndex(), false), actor());
 378
 379         }
 380     }
 381
 382     @Override
 383     public void close() throws Exception {
 384         stopElection();
 385     }
 386
 387     @VisibleForTesting
 388     SnapshotTracker getSnapshotTracker(){
 389         return snapshotTracker;
 390     }
 391
 392     private class InitialSyncStatusTracker {
 393
 394         private static final long INVALID_LOG_INDEX = -2L;
 395         private long initialLeaderCommit = INVALID_LOG_INDEX;
 396         private boolean initialSyncUpDone = false;
 397         private String syncedLeaderId = null;
 398         private final ActorRef actor;
 399
 400         public InitialSyncStatusTracker(ActorRef actor) {
 401             this.actor = actor;
 402         }
 403
 404         public void update(String leaderId, long leaderCommit, long commitIndex){
 405
 406             if(!leaderId.equals(syncedLeaderId)){
 407                 initialSyncUpDone = false;
 408                 initialLeaderCommit = INVALID_LOG_INDEX;
 409                 syncedLeaderId = leaderId;
 410             }
 411
 412             if(!initialSyncUpDone){
 413                 if(initialLeaderCommit == INVALID_LOG_INDEX){
 414                     actor.tell(new FollowerInitialSyncUpStatus(false, getId()), ActorRef.noSender());
 415                     initialLeaderCommit = leaderCommit;
 416                 } else if(commitIndex >= initialLeaderCommit){
 417                     actor.tell(new FollowerInitialSyncUpStatus(true, getId()), ActorRef.noSender());
 418                     initialSyncUpDone = true;
 419                 }
 420             }
 421         }
 422     }
 423 }