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