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 com.google.protobuf.ByteString;
  16 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
  17 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
  18 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
  19 import org.opendaylight.controller.cluster.raft.FollowerLogInformationImpl;
  20 import org.opendaylight.controller.cluster.raft.RaftActorContext;
  21 import org.opendaylight.controller.cluster.raft.RaftState;
  22 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
  23 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
  24 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
  25 import org.opendaylight.controller.cluster.raft.base.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.io.IOException;
  35 import java.util.ArrayList;
  36 import java.util.Collections;
  37 import java.util.HashMap;
  38 import java.util.List;
  39 import java.util.Map;
  40 import java.util.Set;
  41 import java.util.concurrent.TimeUnit;
  42 import java.util.concurrent.atomic.AtomicLong;
  43
  44 /**
  45  * The behavior of a RaftActor when it is in the Leader state
  46  * <p/>
  47  * Leaders:
  48  * <ul>
  49  * <li> Upon election: send initial empty AppendEntries RPCs
  50  * (heartbeat) to each server; repeat during idle periods to
  51  * prevent election timeouts (§5.2)
  52  * <li> If command received from client: append entry to local log,
  53  * respond after entry applied to state machine (§5.3)
  54  * <li> If last log index ≥ nextIndex for a follower: send
  55  * AppendEntries RPC with log entries starting at nextIndex
  56  * <ul>
  57  * <li> If successful: update nextIndex and matchIndex for
  58  * follower (§5.3)
  59  * <li> If AppendEntries fails because of log inconsistency:
  60  * decrement nextIndex and retry (§5.3)
  61  * </ul>
  62  * <li> If there exists an N such that N > commitIndex, a majority
  63  * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
  64  * set commitIndex = N (§5.3, §5.4).
  65  */
  66 public class Leader extends AbstractRaftActorBehavior {
  67
  68
  69     protected final Map<String, FollowerLogInformation> followerToLog =
  70         new HashMap();
  71     protected final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
  72
  73     private final Set<String> followers;
  74
  75     private Cancellable heartbeatSchedule = null;
  76     private Cancellable installSnapshotSchedule = null;
  77
  78     private List<ClientRequestTracker> trackerList = new ArrayList<>();
  79
  80     private final int minReplicationCount;
  81
  82     public Leader(RaftActorContext context) {
  83         super(context);
  84
  85         followers = context.getPeerAddresses().keySet();
  86
  87         for (String followerId : followers) {
  88             FollowerLogInformation followerLogInformation =
  89                 new FollowerLogInformationImpl(followerId,
  90                     new AtomicLong(context.getCommitIndex()),
  91                     new AtomicLong(-1));
  92
  93             followerToLog.put(followerId, followerLogInformation);
  94         }
  95
  96         if(LOG.isDebugEnabled()) {
  97             LOG.debug("Election:Leader has following peers: {}", followers);
  98         }
  99
 100         if (followers.size() > 0) {
 101             minReplicationCount = (followers.size() + 1) / 2 + 1;
 102         } else {
 103             minReplicationCount = 0;
 104         }
 105
 106
 107         // Immediately schedule a heartbeat
 108         // Upon election: send initial empty AppendEntries RPCs
 109         // (heartbeat) to each server; repeat during idle periods to
 110         // prevent election timeouts (§5.2)
 111         scheduleHeartBeat(new FiniteDuration(0, TimeUnit.SECONDS));
 112
 113         scheduleInstallSnapshotCheck(
 114             new FiniteDuration(context.getConfigParams().getHeartBeatInterval().length() * 1000,
 115                 context.getConfigParams().getHeartBeatInterval().unit())
 116         );
 117
 118     }
 119
 120     @Override protected RaftActorBehavior handleAppendEntries(ActorRef sender,
 121         AppendEntries appendEntries) {
 122
 123         if(LOG.isDebugEnabled()) {
 124             LOG.debug(appendEntries.toString());
 125         }
 126
 127         return this;
 128     }
 129
 130     @Override protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
 131         AppendEntriesReply appendEntriesReply) {
 132
 133         if(! appendEntriesReply.isSuccess()) {
 134             if(LOG.isDebugEnabled()) {
 135                 LOG.debug(appendEntriesReply.toString());
 136             }
 137         }
 138
 139         // Update the FollowerLogInformation
 140         String followerId = appendEntriesReply.getFollowerId();
 141         FollowerLogInformation followerLogInformation =
 142             followerToLog.get(followerId);
 143
 144         if(followerLogInformation == null){
 145             LOG.error("Unknown follower {}", followerId);
 146             return this;
 147         }
 148
 149         if (appendEntriesReply.isSuccess()) {
 150             followerLogInformation
 151                 .setMatchIndex(appendEntriesReply.getLogLastIndex());
 152             followerLogInformation
 153                 .setNextIndex(appendEntriesReply.getLogLastIndex() + 1);
 154         } else {
 155
 156             // TODO: When we find that the follower is out of sync with the
 157             // Leader we simply decrement that followers next index by 1.
 158             // Would it be possible to do better than this? The RAFT spec
 159             // does not explicitly deal with it but may be something for us to
 160             // think about
 161
 162             followerLogInformation.decrNextIndex();
 163         }
 164
 165         // Now figure out if this reply warrants a change in the commitIndex
 166         // If there exists an N such that N > commitIndex, a majority
 167         // of matchIndex[i] ≥ N, and log[N].term == currentTerm:
 168         // set commitIndex = N (§5.3, §5.4).
 169         for (long N = context.getCommitIndex() + 1; ; N++) {
 170             int replicatedCount = 1;
 171
 172             for (FollowerLogInformation info : followerToLog.values()) {
 173                 if (info.getMatchIndex().get() >= N) {
 174                     replicatedCount++;
 175                 }
 176             }
 177
 178             if (replicatedCount >= minReplicationCount) {
 179                 ReplicatedLogEntry replicatedLogEntry =
 180                     context.getReplicatedLog().get(N);
 181                 if (replicatedLogEntry != null
 182                     && replicatedLogEntry.getTerm()
 183                     == currentTerm()) {
 184                     context.setCommitIndex(N);
 185                 }
 186             } else {
 187                 break;
 188             }
 189         }
 190
 191         // Apply the change to the state machine
 192         if (context.getCommitIndex() > context.getLastApplied()) {
 193             applyLogToStateMachine(context.getCommitIndex());
 194         }
 195
 196         return this;
 197     }
 198
 199     protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
 200
 201         ClientRequestTracker toRemove = findClientRequestTracker(logIndex);
 202         if(toRemove != null) {
 203             trackerList.remove(toRemove);
 204         }
 205
 206         return toRemove;
 207     }
 208
 209     protected ClientRequestTracker findClientRequestTracker(long logIndex) {
 210         for (ClientRequestTracker tracker : trackerList) {
 211             if (tracker.getIndex() == logIndex) {
 212                 return tracker;
 213             }
 214         }
 215
 216         return null;
 217     }
 218
 219     @Override protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
 220         RequestVoteReply requestVoteReply) {
 221         return this;
 222     }
 223
 224     @Override public RaftState state() {
 225         return RaftState.Leader;
 226     }
 227
 228     @Override public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 229         Preconditions.checkNotNull(sender, "sender should not be null");
 230
 231         Object message = fromSerializableMessage(originalMessage);
 232
 233         if (message instanceof RaftRPC) {
 234             RaftRPC rpc = (RaftRPC) message;
 235             // If RPC request or response contains term T > currentTerm:
 236             // set currentTerm = T, convert to follower (§5.1)
 237             // This applies to all RPC messages and responses
 238             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
 239                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
 240
 241                 return switchBehavior(new Follower(context));
 242             }
 243         }
 244
 245         try {
 246             if (message instanceof SendHeartBeat) {
 247                 sendHeartBeat();
 248                 return this;
 249             } else if(message instanceof SendInstallSnapshot) {
 250                 installSnapshotIfNeeded();
 251             } else if (message instanceof Replicate) {
 252                 replicate((Replicate) message);
 253             } else if (message instanceof InstallSnapshotReply){
 254                 handleInstallSnapshotReply(
 255                     (InstallSnapshotReply) message);
 256             }
 257         } finally {
 258             scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
 259         }
 260
 261         return super.handleMessage(sender, message);
 262     }
 263
 264     private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
 265         String followerId = reply.getFollowerId();
 266         FollowerToSnapshot followerToSnapshot =
 267             mapFollowerToSnapshot.get(followerId);
 268
 269         if (followerToSnapshot != null &&
 270             followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
 271
 272             if (reply.isSuccess()) {
 273                 if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
 274                     //this was the last chunk reply
 275                     if(LOG.isDebugEnabled()) {
 276                         LOG.debug("InstallSnapshotReply received, " +
 277                                 "last chunk received, Chunk:{}. Follower:{} Setting nextIndex:{}",
 278                             reply.getChunkIndex(), followerId,
 279                             context.getReplicatedLog().getSnapshotIndex() + 1
 280                         );
 281                     }
 282
 283                     FollowerLogInformation followerLogInformation =
 284                         followerToLog.get(followerId);
 285                     followerLogInformation.setMatchIndex(
 286                         context.getReplicatedLog().getSnapshotIndex());
 287                     followerLogInformation.setNextIndex(
 288                         context.getReplicatedLog().getSnapshotIndex() + 1);
 289                     mapFollowerToSnapshot.remove(followerId);
 290
 291                     if(LOG.isDebugEnabled()) {
 292                         LOG.debug("followerToLog.get(followerId).getNextIndex().get()=" +
 293                             followerToLog.get(followerId).getNextIndex().get());
 294                     }
 295
 296                 } else {
 297                     followerToSnapshot.markSendStatus(true);
 298                 }
 299             } else {
 300                 LOG.info("InstallSnapshotReply received, " +
 301                         "sending snapshot chunk failed, Will retry, Chunk:{}",
 302                     reply.getChunkIndex()
 303                 );
 304                 followerToSnapshot.markSendStatus(false);
 305             }
 306
 307         } else {
 308             LOG.error("ERROR!!" +
 309                     "FollowerId in InstallSnapshotReply not known to Leader" +
 310                     " or Chunk Index in InstallSnapshotReply not matching {} != {}",
 311                 followerToSnapshot.getChunkIndex(), reply.getChunkIndex()
 312             );
 313         }
 314     }
 315
 316     private void replicate(Replicate replicate) {
 317         long logIndex = replicate.getReplicatedLogEntry().getIndex();
 318
 319         if(LOG.isDebugEnabled()) {
 320             LOG.debug("Replicate message {}", logIndex);
 321         }
 322
 323         // Create a tracker entry we will use this later to notify the
 324         // client actor
 325         trackerList.add(
 326             new ClientRequestTrackerImpl(replicate.getClientActor(),
 327                 replicate.getIdentifier(),
 328                 logIndex)
 329         );
 330
 331         if (followers.size() == 0) {
 332             context.setCommitIndex(logIndex);
 333             applyLogToStateMachine(logIndex);
 334         } else {
 335             sendAppendEntries();
 336         }
 337     }
 338
 339     private void sendAppendEntries() {
 340         // Send an AppendEntries to all followers
 341         for (String followerId : followers) {
 342             ActorSelection followerActor = context.getPeerActorSelection(followerId);
 343
 344             if (followerActor != null) {
 345                 FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
 346                 long followerNextIndex = followerLogInformation.getNextIndex().get();
 347                 List<ReplicatedLogEntry> entries = Collections.emptyList();
 348
 349                 if (mapFollowerToSnapshot.get(followerId) != null) {
 350                     if (mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
 351                         sendSnapshotChunk(followerActor, followerId);
 352                     }
 353
 354                 } else {
 355
 356                     if (context.getReplicatedLog().isPresent(followerNextIndex)) {
 357                         // FIXME : Sending one entry at a time
 358                         entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
 359
 360                         followerActor.tell(
 361                             new AppendEntries(currentTerm(), context.getId(),
 362                                 prevLogIndex(followerNextIndex),
 363                                 prevLogTerm(followerNextIndex), entries,
 364                                 context.getCommitIndex()).toSerializable(),
 365                             actor()
 366                         );
 367
 368                     } else {
 369                         // if the followers next index is not present in the leaders log, then snapshot should be sent
 370                         long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 371                         long leaderLastIndex = context.getReplicatedLog().lastIndex();
 372                         if (followerNextIndex >= 0 && leaderLastIndex >= followerNextIndex ) {
 373                             // if the follower is just not starting and leader's index
 374                             // is more than followers index
 375                             if(LOG.isDebugEnabled()) {
 376                                 LOG.debug("SendInstallSnapshot to follower:{}," +
 377                                         "follower-nextIndex:{}, leader-snapshot-index:{},  " +
 378                                         "leader-last-index:{}", followerId,
 379                                     followerNextIndex, leaderSnapShotIndex, leaderLastIndex
 380                                 );
 381                             }
 382
 383                             actor().tell(new SendInstallSnapshot(), actor());
 384                         } else {
 385                             followerActor.tell(
 386                                 new AppendEntries(currentTerm(), context.getId(),
 387                                     prevLogIndex(followerNextIndex),
 388                                     prevLogTerm(followerNextIndex), entries,
 389                                     context.getCommitIndex()).toSerializable(),
 390                                 actor()
 391                             );
 392                         }
 393                     }
 394                 }
 395             }
 396         }
 397     }
 398
 399     /**
 400      * An installSnapshot is scheduled at a interval that is a multiple of
 401      * a HEARTBEAT_INTERVAL. This is to avoid the need to check for installing
 402      * snapshots at every heartbeat.
 403      */
 404     private void installSnapshotIfNeeded(){
 405         for (String followerId : followers) {
 406             ActorSelection followerActor =
 407                 context.getPeerActorSelection(followerId);
 408
 409             if(followerActor != null) {
 410                 FollowerLogInformation followerLogInformation =
 411                     followerToLog.get(followerId);
 412
 413                 long nextIndex = followerLogInformation.getNextIndex().get();
 414
 415                 if (!context.getReplicatedLog().isPresent(nextIndex) &&
 416                     context.getReplicatedLog().isInSnapshot(nextIndex)) {
 417                     sendSnapshotChunk(followerActor, followerId);
 418                 }
 419             }
 420         }
 421     }
 422
 423     /**
 424      *  Sends a snapshot chunk to a given follower
 425      *  InstallSnapshot should qualify as a heartbeat too.
 426      */
 427     private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
 428         try {
 429             followerActor.tell(
 430                 new InstallSnapshot(currentTerm(), context.getId(),
 431                     context.getReplicatedLog().getSnapshotIndex(),
 432                     context.getReplicatedLog().getSnapshotTerm(),
 433                     getNextSnapshotChunk(followerId,
 434                         context.getReplicatedLog().getSnapshot()),
 435                     mapFollowerToSnapshot.get(followerId).incrementChunkIndex(),
 436                     mapFollowerToSnapshot.get(followerId).getTotalChunks()
 437                 ).toSerializable(),
 438                 actor()
 439             );
 440             LOG.info("InstallSnapshot sent to follower {}, Chunk: {}/{}",
 441                 followerActor.path(), mapFollowerToSnapshot.get(followerId).getChunkIndex(),
 442                 mapFollowerToSnapshot.get(followerId).getTotalChunks());
 443         } catch (IOException e) {
 444             LOG.error(e, "InstallSnapshot failed for Leader.");
 445         }
 446     }
 447
 448     /**
 449      * Acccepts snaphot as ByteString, enters into map for future chunks
 450      * creates and return a ByteString chunk
 451      */
 452     private ByteString getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException {
 453         FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId);
 454         if (followerToSnapshot == null) {
 455             followerToSnapshot = new FollowerToSnapshot(snapshotBytes);
 456             mapFollowerToSnapshot.put(followerId, followerToSnapshot);
 457         }
 458         ByteString nextChunk = followerToSnapshot.getNextChunk();
 459         if(LOG.isDebugEnabled()) {
 460             LOG.debug("Leader's snapshot nextChunk size:{}", nextChunk.size());
 461         }
 462
 463         return nextChunk;
 464     }
 465
 466     private void sendHeartBeat() {
 467         if (followers.size() > 0) {
 468             sendAppendEntries();
 469         }
 470     }
 471
 472     private void stopHeartBeat() {
 473         if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
 474             heartbeatSchedule.cancel();
 475         }
 476     }
 477
 478     private void stopInstallSnapshotSchedule() {
 479         if (installSnapshotSchedule != null && !installSnapshotSchedule.isCancelled()) {
 480             installSnapshotSchedule.cancel();
 481         }
 482     }
 483
 484     private void scheduleHeartBeat(FiniteDuration interval) {
 485         if(followers.size() == 0){
 486             // Optimization - do not bother scheduling a heartbeat as there are
 487             // no followers
 488             return;
 489         }
 490
 491         stopHeartBeat();
 492
 493         // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
 494         // message is sent to itself.
 495         // Scheduling the heartbeat only once here because heartbeats do not
 496         // need to be sent if there are other messages being sent to the remote
 497         // actor.
 498         heartbeatSchedule =
 499             context.getActorSystem().scheduler().scheduleOnce(
 500                 interval,
 501                 context.getActor(), new SendHeartBeat(),
 502                 context.getActorSystem().dispatcher(), context.getActor());
 503     }
 504
 505
 506     private void scheduleInstallSnapshotCheck(FiniteDuration interval) {
 507         if(followers.size() == 0){
 508             // Optimization - do not bother scheduling a heartbeat as there are
 509             // no followers
 510             return;
 511         }
 512
 513         stopInstallSnapshotSchedule();
 514
 515         // Schedule a message to send append entries to followers that can
 516         // accept an append entries with some data in it
 517         installSnapshotSchedule =
 518             context.getActorSystem().scheduler().scheduleOnce(
 519                 interval,
 520                 context.getActor(), new SendInstallSnapshot(),
 521                 context.getActorSystem().dispatcher(), context.getActor());
 522     }
 523
 524
 525
 526     @Override public void close() throws Exception {
 527         stopHeartBeat();
 528     }
 529
 530     @Override public String getLeaderId() {
 531         return context.getId();
 532     }
 533
 534     /**
 535      * Encapsulates the snapshot bytestring and handles the logic of sending
 536      * snapshot chunks
 537      */
 538     protected class FollowerToSnapshot {
 539         private ByteString snapshotBytes;
 540         private int offset = 0;
 541         // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset
 542         private int replyReceivedForOffset;
 543         // if replyStatus is false, the previous chunk is attempted
 544         private boolean replyStatus = false;
 545         private int chunkIndex;
 546         private int totalChunks;
 547
 548         public FollowerToSnapshot(ByteString snapshotBytes) {
 549             this.snapshotBytes = snapshotBytes;
 550             replyReceivedForOffset = -1;
 551             chunkIndex = 1;
 552             int size = snapshotBytes.size();
 553             totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) +
 554                 ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0);
 555             if(LOG.isDebugEnabled()) {
 556                 LOG.debug("Snapshot {} bytes, total chunks to send:{}",
 557                     size, totalChunks);
 558             }
 559         }
 560
 561         public ByteString getSnapshotBytes() {
 562             return snapshotBytes;
 563         }
 564
 565         public int incrementOffset() {
 566             if(replyStatus) {
 567                 // if prev chunk failed, we would want to sent the same chunk again
 568                 offset = offset + context.getConfigParams().getSnapshotChunkSize();
 569             }
 570             return offset;
 571         }
 572
 573         public int incrementChunkIndex() {
 574             if (replyStatus) {
 575                 // if prev chunk failed, we would want to sent the same chunk again
 576                 chunkIndex =  chunkIndex + 1;
 577             }
 578             return chunkIndex;
 579         }
 580
 581         public int getChunkIndex() {
 582             return chunkIndex;
 583         }
 584
 585         public int getTotalChunks() {
 586             return totalChunks;
 587         }
 588
 589         public boolean canSendNextChunk() {
 590             // we only send a false if a chunk is sent but we have not received a reply yet
 591             return replyReceivedForOffset == offset;
 592         }
 593
 594         public boolean isLastChunk(int chunkIndex) {
 595             return totalChunks == chunkIndex;
 596         }
 597
 598         public void markSendStatus(boolean success) {
 599             if (success) {
 600                 // if the chunk sent was successful
 601                 replyReceivedForOffset = offset;
 602                 replyStatus = true;
 603             } else {
 604                 // if the chunk sent was failure
 605                 replyReceivedForOffset = offset;
 606                 replyStatus = false;
 607             }
 608         }
 609
 610         public ByteString getNextChunk() {
 611             int snapshotLength = getSnapshotBytes().size();
 612             int start = incrementOffset();
 613             int size = context.getConfigParams().getSnapshotChunkSize();
 614             if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) {
 615                 size = snapshotLength;
 616             } else {
 617                 if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) {
 618                     size = snapshotLength - start;
 619                 }
 620             }
 621
 622             if(LOG.isDebugEnabled()) {
 623                 LOG.debug("length={}, offset={},size={}",
 624                     snapshotLength, start, size);
 625             }
 626             return getSnapshotBytes().substring(start, start + size);
 627
 628         }
 629     }
 630
 631 }