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