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