opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/Leader.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 akka.actor.ActorSelection;
  13 import akka.actor.Cancellable;
  14 import com.google.common.base.Preconditions;
  15 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
  16 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
  17 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
  18 import org.opendaylight.controller.cluster.raft.FollowerLogInformationImpl;
  19 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  20 import org.opendaylight.controller.cluster.raft.RaftState;
  21 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  22 import org.opendaylight.controller.cluster.raft.internal.messages.ApplyState;
  23 import org.opendaylight.controller.cluster.raft.internal.messages.Replicate;
  24 import org.opendaylight.controller.cluster.raft.internal.messages.SendHeartBeat;
  25 import org.opendaylight.controller.cluster.raft.internal.messages.SendInstallSnapshot;
  26 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
  27 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  28 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
  29 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
  30 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
  31 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  32 import scala.concurrent.duration.FiniteDuration;
  33
  34 import java.util.ArrayList;
  35 import java.util.Collections;
  36 import java.util.HashMap;
  37 import java.util.List;
  38 import java.util.Map;
  39 import java.util.concurrent.TimeUnit;
  40 import java.util.concurrent.atomic.AtomicLong;
  41
  42 /**
  43  * The behavior of a RaftActor when it is in the Leader state
  44  * <p/>
  45  * Leaders:
  46  * <ul>
  47  * <li> Upon election: send initial empty AppendEntries RPCs
  48  * (heartbeat) to each server; repeat during idle periods to
  49  * prevent election timeouts (§5.2)
  50  * <li> If command received from client: append entry to local log,
  51  * respond after entry applied to state machine (§5.3)
  52  * <li> If last log index ≥ nextIndex for a follower: send
  53  * AppendEntries RPC with log entries starting at nextIndex
  54  * <ul>
  55  * <li> If successful: update nextIndex and matchIndex for
  56  * follower (§5.3)
  57  * <li> If AppendEntries fails because of log inconsistency:
  58  * decrement nextIndex and retry (§5.3)
  59  * </ul>
  60  * <li> If there exists an N such that N > commitIndex, a majority
  61  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  62  * set commitIndex = N (§5.3, §5.4).
  63  */
  64 public class Leader extends AbstractRaftActorBehavior {
  65
  66
  67     private final Map<String, FollowerLogInformation> followerToLog =
  68         new HashMap();
  69
  70     private final Map<String, ActorSelection> followerToActor = new HashMap<>();
  71
  72     private Cancellable heartbeatSchedule = null;
  73     private Cancellable appendEntriesSchedule = null;
  74     private Cancellable installSnapshotSchedule = null;
  75
  76     private List<ClientRequestTracker> trackerList = new ArrayList<>();
  77
  78     private final int minReplicationCount;
  79
  80     public Leader(RaftActorContext context) {
  81         super(context);
  82
  83         if (lastIndex() >= 0) {
  84             context.setCommitIndex(lastIndex());
  85         }
  86
  87         for (String followerId : context.getPeerAddresses().keySet()) {
  88             FollowerLogInformation followerLogInformation =
  89                 new FollowerLogInformationImpl(followerId,
  90                     new AtomicLong(lastIndex()),
  91                     new AtomicLong(-1));
  92
  93             followerToActor.put(followerId,
  94                 context.actorSelection(context.getPeerAddress(followerId)));
  95
  96             followerToLog.put(followerId, followerLogInformation);
  97         }
  98
  99         context.getLogger().debug("Election:Leader has following peers:"+followerToActor.keySet());
 100
 101         if (followerToActor.size() > 0) {
 102             minReplicationCount = (followerToActor.size() + 1) / 2 + 1;
 103         } else {
 104             minReplicationCount = 0;
 105         }
 106
 107
 108         // Immediately schedule a heartbeat
 109         // Upon election: send initial empty AppendEntries RPCs
 110         // (heartbeat) to each server; repeat during idle periods to
 111         // prevent election timeouts (§5.2)
 112         scheduleHeartBeat(new FiniteDuration(0, TimeUnit.SECONDS));
 113
 114         scheduleInstallSnapshotCheck(
 115             new FiniteDuration(HEART_BEAT_INTERVAL.length() * 1000,
 116                 HEART_BEAT_INTERVAL.unit())
 117         );
 118
 119     }
 120
 121     @Override protected RaftState handleAppendEntries(ActorRef sender,
 122         AppendEntries appendEntries) {
 123
 124         return state();
 125     }
 126
 127     @Override protected RaftState handleAppendEntriesReply(ActorRef sender,
 128         AppendEntriesReply appendEntriesReply) {
 129
 130         // Update the FollowerLogInformation
 131         String followerId = appendEntriesReply.getFollowerId();
 132         FollowerLogInformation followerLogInformation =
 133             followerToLog.get(followerId);
 134         if (appendEntriesReply.isSuccess()) {
 135             followerLogInformation
 136                 .setMatchIndex(appendEntriesReply.getLogLastIndex());
 137             followerLogInformation
 138                 .setNextIndex(appendEntriesReply.getLogLastIndex() + 1);
 139         } else {
 140
 141             // TODO: When we find that the follower is out of sync with the
 142             // Leader we simply decrement that followers next index by 1.
 143             // Would it be possible to do better than this? The RAFT spec
 144             // does not explicitly deal with it but may be something for us to
 145             // think about
 146
 147             followerLogInformation.decrNextIndex();
 148         }
 149
 150         // Now figure out if this reply warrants a change in the commitIndex
 151         // If there exists an N such that N > commitIndex, a majority
 152         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 153         // set commitIndex = N (§5.3, §5.4).
 154         for (long N = context.getCommitIndex() + 1; ; N++) {
 155             int replicatedCount = 1;
 156
 157             for (FollowerLogInformation info : followerToLog.values()) {
 158                 if (info.getMatchIndex().get() >= N) {
 159                     replicatedCount++;
 160                 }
 161             }
 162
 163             if (replicatedCount >= minReplicationCount) {
 164                 ReplicatedLogEntry replicatedLogEntry =
 165                     context.getReplicatedLog().get(N);
 166                 if (replicatedLogEntry != null
 167                     && replicatedLogEntry.getTerm()
 168                     == currentTerm()) {
 169                     context.setCommitIndex(N);
 170                 }
 171             } else {
 172                 break;
 173             }
 174         }
 175
 176         // Apply the change to the state machine
 177         if (context.getCommitIndex() > context.getLastApplied()) {
 178             applyLogToStateMachine(context.getCommitIndex());
 179         }
 180
 181         return state();
 182     }
 183
 184     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 185         for (ClientRequestTracker tracker : trackerList) {
 186             if (tracker.getIndex() == logIndex) {
 187                 return tracker;
 188             }
 189         }
 190
 191         return null;
 192     }
 193
 194     @Override protected RaftState handleRequestVoteReply(ActorRef sender,
 195         RequestVoteReply requestVoteReply) {
 196         return state();
 197     }
 198
 199     @Override public RaftState state() {
 200         return RaftState.Leader;
 201     }
 202
 203     @Override public RaftState handleMessage(ActorRef sender, Object originalMessage) {
 204         Preconditions.checkNotNull(sender, "sender should not be null");
 205
 206         Object message = fromSerializableMessage(originalMessage);
 207
 208         if (message instanceof RaftRPC) {
 209             RaftRPC rpc = (RaftRPC) message;
 210             // If RPC request or response contains term T > currentTerm:
 211             // set currentTerm = T, convert to follower (§5.1)
 212             // This applies to all RPC messages and responses
 213             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 214                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 215                 return RaftState.Follower;
 216             }
 217         }
 218
 219         try {
 220             if (message instanceof SendHeartBeat) {
 221                 return sendHeartBeat();
 222             } else if(message instanceof SendInstallSnapshot) {
 223                 installSnapshotIfNeeded();
 224             } else if (message instanceof Replicate) {
 225                 replicate((Replicate) message);
 226             } else if (message instanceof InstallSnapshotReply){
 227                 handleInstallSnapshotReply(
 228                     (InstallSnapshotReply) message);
 229             }
 230         } finally {
 231             scheduleHeartBeat(HEART_BEAT_INTERVAL);
 232         }
 233
 234         return super.handleMessage(sender, message);
 235     }
 236
 237     private void handleInstallSnapshotReply(InstallSnapshotReply message) {
 238         InstallSnapshotReply reply = message;
 239         String followerId = reply.getFollowerId();
 240         FollowerLogInformation followerLogInformation =
 241             followerToLog.get(followerId);
 242
 243         followerLogInformation
 244             .setMatchIndex(context.getReplicatedLog().getSnapshotIndex());
 245         followerLogInformation
 246             .setNextIndex(context.getReplicatedLog().getSnapshotIndex() + 1);
 247     }
 248
 249     private void replicate(Replicate replicate) {
 250         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 251
 252         context.getLogger().debug("Replicate message " + logIndex);
 253
 254         if (followerToActor.size() == 0) {
 255             context.setCommitIndex(
 256                 replicate.getReplicatedLogEntry().getIndex());
 257
 258             context.getActor()
 259                 .tell(new ApplyState(replicate.getClientActor(),
 260                         replicate.getIdentifier(),
 261                         replicate.getReplicatedLogEntry()),
 262                     context.getActor()
 263                 );
 264         } else {
 265
 266             // Create a tracker entry we will use this later to notify the
 267             // client actor
 268             trackerList.add(
 269                 new ClientRequestTrackerImpl(replicate.getClientActor(),
 270                     replicate.getIdentifier(),
 271                     logIndex)
 272             );
 273
 274             sendAppendEntries();
 275         }
 276     }
 277
 278     private void sendAppendEntries() {
 279         // Send an AppendEntries to all followers
 280         for (String followerId : followerToActor.keySet()) {
 281             ActorSelection followerActor =
 282                 followerToActor.get(followerId);
 283
 284             FollowerLogInformation followerLogInformation =
 285                 followerToLog.get(followerId);
 286
 287             long nextIndex = followerLogInformation.getNextIndex().get();
 288
 289             List<ReplicatedLogEntry> entries = Collections.emptyList();
 290
 291             if(context.getReplicatedLog().isPresent(nextIndex)){
 292                 // TODO: Instead of sending all entries from nextIndex
 293                 // only send a fixed number of entries to each follower
 294                 // This is to avoid the situation where there are a lot of
 295                 // entries to install for a fresh follower or to a follower
 296                 // that has fallen too far behind with the log but yet is not
 297                 // eligible to receive a snapshot
 298                 entries =
 299                     context.getReplicatedLog().getFrom(nextIndex);
 300             }
 301
 302             followerActor.tell(
 303                 new AppendEntries(currentTerm(), context.getId(), prevLogIndex(nextIndex),
 304                     prevLogTerm(nextIndex), entries, context.getCommitIndex()).toSerializable(),
 305                 actor());
 306         }
 307     }
 308
 309     /**
 310      * An installSnapshot is scheduled at a interval that is a multiple of
 311      * a HEARTBEAT_INTERVAL. This is to avoid the need to check for installing
 312      * snapshots at every heartbeat.
 313      */
 314     private void installSnapshotIfNeeded(){
 315         for (String followerId : followerToActor.keySet()) {
 316             ActorSelection followerActor =
 317                 followerToActor.get(followerId);
 318
 319             FollowerLogInformation followerLogInformation =
 320                 followerToLog.get(followerId);
 321
 322             long nextIndex = followerLogInformation.getNextIndex().get();
 323
 324             if(!context.getReplicatedLog().isPresent(nextIndex) && context.getReplicatedLog().isInSnapshot(nextIndex)){
 325                 followerActor.tell(
 326                     new InstallSnapshot(currentTerm(), context.getId(),
 327                         context.getReplicatedLog().getSnapshotIndex(),
 328                         context.getReplicatedLog().getSnapshotTerm(),
 329                         context.getReplicatedLog().getSnapshot()
 330                     ),
 331                     actor()
 332                 );
 333             }
 334         }
 335     }
 336
 337     private RaftState sendHeartBeat() {
 338         if (followerToActor.size() > 0) {
 339             sendAppendEntries();
 340         }
 341         return state();
 342     }
 343
 344     private void stopHeartBeat() {
 345         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 346             heartbeatSchedule.cancel();
 347         }
 348     }
 349
 350     private void stopInstallSnapshotSchedule() {
 351         if (installSnapshotSchedule != null && !installSnapshotSchedule.isCancelled()) {
 352             installSnapshotSchedule.cancel();
 353         }
 354     }
 355
 356     private void scheduleHeartBeat(FiniteDuration interval) {
 357         if(followerToActor.keySet().size() == 0){
 358             // Optimization - do not bother scheduling a heartbeat as there are
 359             // no followers
 360             return;
 361         }
 362
 363         stopHeartBeat();
 364
 365         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 366         // message is sent to itself.
 367         // Scheduling the heartbeat only once here because heartbeats do not
 368         // need to be sent if there are other messages being sent to the remote
 369         // actor.
 370         heartbeatSchedule =
 371             context.getActorSystem().scheduler().scheduleOnce(
 372                 interval,
 373                 context.getActor(), new SendHeartBeat(),
 374                 context.getActorSystem().dispatcher(), context.getActor());
 375     }
 376
 377
 378     private void scheduleInstallSnapshotCheck(FiniteDuration interval) {
 379         if(followerToActor.keySet().size() == 0){
 380             // Optimization - do not bother scheduling a heartbeat as there are
 381             // no followers
 382             return;
 383         }
 384
 385         stopInstallSnapshotSchedule();
 386
 387         // Schedule a message to send append entries to followers that can
 388         // accept an append entries with some data in it
 389         installSnapshotSchedule =
 390             context.getActorSystem().scheduler().scheduleOnce(
 391                 interval,
 392                 context.getActor(), new SendInstallSnapshot(),
 393                 context.getActorSystem().dispatcher(), context.getActor());
 394     }
 395
 396
 397
 398     @Override public void close() throws Exception {
 399         stopHeartBeat();
 400     }
 401
 402     @Override public String getLeaderId() {
 403         return context.getId();
 404     }
 405
 406 }