2 * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved.
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
9 package org.opendaylight.controller.cluster.raft.behaviors;
11 import akka.actor.ActorRef;
12 import akka.actor.ActorSelection;
13 import akka.actor.Address;
14 import akka.cluster.Cluster;
15 import akka.cluster.ClusterEvent.CurrentClusterState;
16 import akka.cluster.Member;
17 import akka.cluster.MemberStatus;
18 import akka.japi.Procedure;
19 import com.google.common.annotations.VisibleForTesting;
20 import com.google.common.base.Stopwatch;
21 import java.util.ArrayList;
22 import java.util.Optional;
24 import java.util.concurrent.TimeUnit;
25 import javax.annotation.Nullable;
26 import org.opendaylight.controller.cluster.raft.RaftActorContext;
27 import org.opendaylight.controller.cluster.raft.RaftState;
28 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
29 import org.opendaylight.controller.cluster.raft.Snapshot;
30 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
31 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
32 import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow;
33 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
34 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
35 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
36 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
37 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
38 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
39 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
40 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
43 * The behavior of a RaftActor in the Follower state
46 * <li> Respond to RPCs from candidates and leaders
47 * <li> If election timeout elapses without receiving AppendEntries
48 * RPC from current leader or granting vote to candidate:
49 * convert to candidate
52 public class Follower extends AbstractRaftActorBehavior {
53 private static final int SYNC_THRESHOLD = 10;
55 private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
57 private final SyncStatusTracker initialSyncStatusTracker;
59 private final Procedure<ReplicatedLogEntry> appendAndPersistCallback = new Procedure<ReplicatedLogEntry>() {
61 public void apply(ReplicatedLogEntry logEntry) {
62 context.getReplicatedLog().captureSnapshotIfReady(logEntry);
66 private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted();
67 private SnapshotTracker snapshotTracker = null;
68 private String leaderId;
69 private short leaderPayloadVersion;
71 public Follower(RaftActorContext context) {
72 this(context, null, (short)-1);
75 public Follower(RaftActorContext context, String initialLeaderId, short initialLeaderPayloadVersion) {
76 super(context, RaftState.Follower);
77 this.leaderId = initialLeaderId;
78 this.leaderPayloadVersion = initialLeaderPayloadVersion;
80 initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), SYNC_THRESHOLD);
82 if (context.getPeerIds().isEmpty() && getLeaderId() == null) {
83 actor().tell(TimeoutNow.INSTANCE, actor());
85 scheduleElection(electionDuration());
90 public final String getLeaderId() {
95 protected final void setLeaderId(@Nullable final String leaderId) {
96 this.leaderId = leaderId;
100 public short getLeaderPayloadVersion() {
101 return leaderPayloadVersion;
105 protected final void setLeaderPayloadVersion(short leaderPayloadVersion) {
106 this.leaderPayloadVersion = leaderPayloadVersion;
109 private void restartLastLeaderMessageTimer() {
110 if (lastLeaderMessageTimer.isRunning()) {
111 lastLeaderMessageTimer.reset();
114 lastLeaderMessageTimer.start();
117 private boolean isLogEntryPresent(long index){
118 if(context.getReplicatedLog().isInSnapshot(index)) {
122 ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
123 return entry != null;
127 private void updateInitialSyncStatus(long currentLeaderCommit, String leaderId){
128 initialSyncStatusTracker.update(leaderId, currentLeaderCommit, context.getCommitIndex());
132 protected RaftActorBehavior handleAppendEntries(ActorRef sender, AppendEntries appendEntries) {
134 int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
135 if(LOG.isTraceEnabled()) {
136 LOG.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
137 } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
138 LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
141 // TODO : Refactor this method into a bunch of smaller methods
142 // to make it easier to read. Before refactoring ensure tests
143 // cover the code properly
145 if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
146 // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
147 AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
148 lastIndex(), lastTerm(), context.getPayloadVersion());
150 if(LOG.isDebugEnabled()) {
151 LOG.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
153 sender.tell(reply, actor());
158 // If we got here then we do appear to be talking to the leader
159 leaderId = appendEntries.getLeaderId();
160 leaderPayloadVersion = appendEntries.getPayloadVersion();
162 updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
163 // First check if the logs are in sync or not
164 long lastIndex = lastIndex();
166 if (isOutOfSync(appendEntries)) {
167 // We found that the log was out of sync so just send a negative
170 LOG.debug("{}: Follower is out-of-sync, so sending negative reply, lastIndex: {}, lastTerm: {}",
171 logName(), lastIndex, lastTerm());
173 sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
174 lastTerm(), context.getPayloadVersion()), actor());
178 if (appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
180 LOG.debug("{}: Number of entries to be appended = {}", logName(),
181 appendEntries.getEntries().size());
183 // 3. If an existing entry conflicts with a new one (same index
184 // but different terms), delete the existing entry and all that
186 int addEntriesFrom = 0;
187 if (context.getReplicatedLog().size() > 0) {
189 // Find the entry up until the one that is not in the follower's log
190 for (int i = 0;i < appendEntries.getEntries().size(); i++, addEntriesFrom++) {
191 ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
193 if(!isLogEntryPresent(matchEntry.getIndex())) {
194 // newEntry not found in the log
198 long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex());
200 LOG.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry,
203 // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know
204 // what the term was so we'll assume it matches.
205 if(existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) {
209 if(!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
211 LOG.debug("{}: Removing entries from log starting at {}", logName(),
212 matchEntry.getIndex());
214 // Entries do not match so remove all subsequent entries
215 if(!context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) {
216 // Could not remove the entries - this means the matchEntry index must be in the
217 // snapshot and not the log. In this case the prior entries are part of the state
218 // so we must send back a reply to force a snapshot to completely re-sync the
219 // follower's log and state.
221 LOG.debug("{}: Could not remove entries - sending reply to force snapshot", logName());
222 sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
223 lastTerm(), context.getPayloadVersion(), true), actor());
229 sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
230 lastTerm(), context.getPayloadVersion(), true), actor());
236 lastIndex = lastIndex();
237 LOG.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(),
238 lastIndex, addEntriesFrom);
240 // 4. Append any new entries not already in the log
241 for (int i = addEntriesFrom; i < appendEntries.getEntries().size(); i++) {
242 ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
244 LOG.debug("{}: Append entry to log {}", logName(), entry.getData());
246 context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback);
248 if(entry.getData() instanceof ServerConfigurationPayload) {
249 context.updatePeerIds((ServerConfigurationPayload)entry.getData());
253 LOG.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
256 // 5. If leaderCommit > commitIndex, set commitIndex =
257 // min(leaderCommit, index of last new entry)
259 lastIndex = lastIndex();
260 long prevCommitIndex = context.getCommitIndex();
262 context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
264 if (prevCommitIndex != context.getCommitIndex()) {
265 LOG.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
268 // If commitIndex > lastApplied: increment lastApplied, apply
269 // log[lastApplied] to state machine (§5.3)
270 // check if there are any entries to be applied. last-applied can be equal to last-index
271 if (appendEntries.getLeaderCommit() > context.getLastApplied() &&
272 context.getLastApplied() < lastIndex) {
273 if(LOG.isDebugEnabled()) {
274 LOG.debug("{}: applyLogToStateMachine, " +
275 "appendEntries.getLeaderCommit(): {}," +
276 "context.getLastApplied(): {}, lastIndex(): {}", logName(),
277 appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
280 applyLogToStateMachine(appendEntries.getLeaderCommit());
283 AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
284 lastIndex, lastTerm(), context.getPayloadVersion());
286 if(LOG.isTraceEnabled()) {
287 LOG.trace("{}: handleAppendEntries returning : {}", logName(), reply);
288 } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
289 LOG.debug("{}: handleAppendEntries returning : {}", logName(), reply);
292 sender.tell(reply, actor());
294 if (!context.getSnapshotManager().isCapturing()) {
295 super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
301 private boolean isOutOfSync(AppendEntries appendEntries) {
303 long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex());
304 boolean prevEntryPresent = isLogEntryPresent(appendEntries.getPrevLogIndex());
305 long lastIndex = lastIndex();
306 int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
307 boolean outOfSync = true;
309 if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
311 // The follower's log is out of sync because the leader does have
312 // an entry at prevLogIndex and this follower has no entries in
315 LOG.debug("{}: The followers log is empty and the senders prevLogIndex is {}",
316 logName(), appendEntries.getPrevLogIndex());
317 } else if (lastIndex > -1 && appendEntries.getPrevLogIndex() != -1 && !prevEntryPresent) {
319 // The follower's log is out of sync because the Leader's
320 // prevLogIndex entry was not found in it's log
322 LOG.debug("{}: The log is not empty but the prevLogIndex {} was not found in it - lastIndex: {}, snapshotIndex: {}",
323 logName(), appendEntries.getPrevLogIndex(), lastIndex, context.getReplicatedLog().getSnapshotIndex());
324 } else if (lastIndex > -1 && prevEntryPresent && prevLogTerm != appendEntries.getPrevLogTerm()) {
326 // The follower's log is out of sync because the Leader's
327 // prevLogIndex entry does exist in the follower's log but it has
328 // a different term in it
331 "{}: Cannot append entries because previous entry term {} is not equal to append entries prevLogTerm {}",
332 logName(), prevLogTerm, appendEntries.getPrevLogTerm());
333 } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
334 && appendEntries.getReplicatedToAllIndex() != -1
335 && !isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
336 // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
337 // the previous entry in it's in-memory journal
340 "{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the in-memory journal",
341 logName(), appendEntries.getReplicatedToAllIndex());
342 } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
343 && appendEntries.getReplicatedToAllIndex() != -1 && numLogEntries > 0
344 && !isLogEntryPresent(appendEntries.getEntries().get(0).getIndex() - 1)) {
346 "{}: Cannot append entries because the calculated previousIndex {} was not found in the in-memory journal",
347 logName(), appendEntries.getEntries().get(0).getIndex() - 1);
355 protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
356 AppendEntriesReply appendEntriesReply) {
361 protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
362 RequestVoteReply requestVoteReply) {
367 public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
368 if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
369 return handleElectionTimeout(message);
372 if (!(message instanceof RaftRPC)) {
373 // The rest of the processing requires the message to be a RaftRPC
377 final RaftRPC rpc = (RaftRPC) message;
378 // If RPC request or response contains term T > currentTerm:
379 // set currentTerm = T, convert to follower (§5.1)
380 // This applies to all RPC messages and responses
381 if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
382 LOG.debug("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
383 logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
385 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
388 if (rpc instanceof InstallSnapshot) {
389 handleInstallSnapshot(sender, (InstallSnapshot) rpc);
390 restartLastLeaderMessageTimer();
391 scheduleElection(electionDuration());
395 if (!(rpc instanceof RequestVote) || canGrantVote((RequestVote) rpc)) {
396 restartLastLeaderMessageTimer();
397 scheduleElection(electionDuration());
400 return super.handleMessage(sender, rpc);
403 private RaftActorBehavior handleElectionTimeout(Object message) {
404 // If the message is ElectionTimeout, verify we haven't actually seen a message from the leader
405 // during the election timeout interval. It may that the election timer expired b/c this actor
406 // was busy and messages got delayed, in which case leader messages would be backed up in the
407 // queue but would be processed before the ElectionTimeout message and thus would restart the
408 // lastLeaderMessageTimer.
409 long lastLeaderMessageInterval = lastLeaderMessageTimer.elapsed(TimeUnit.MILLISECONDS);
410 long electionTimeoutInMillis = context.getConfigParams().getElectionTimeOutInterval().toMillis();
411 boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning() ||
412 lastLeaderMessageInterval >= electionTimeoutInMillis;
414 if(canStartElection()) {
415 if(message instanceof TimeoutNow) {
416 LOG.debug("{}: Received TimeoutNow - switching to Candidate", logName());
417 return internalSwitchBehavior(RaftState.Candidate);
418 } else if(noLeaderMessageReceived) {
419 // Check the cluster state to see if the leader is known to be up before we go to Candidate.
420 // However if we haven't heard from the leader in a long time even though the cluster state
421 // indicates it's up then something is wrong - leader might be stuck indefinitely - so switch
423 long maxElectionTimeout = electionTimeoutInMillis * MAX_ELECTION_TIMEOUT_FACTOR;
424 if(isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
425 LOG.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
426 scheduleElection(electionDuration());
428 LOG.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
429 return internalSwitchBehavior(RaftState.Candidate);
432 LOG.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
433 logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
434 scheduleElection(electionDuration());
436 } else if(message instanceof ElectionTimeout) {
437 if(noLeaderMessageReceived) {
441 scheduleElection(electionDuration());
447 private boolean isLeaderAvailabilityKnown() {
448 if(leaderId == null) {
452 Optional<Cluster> cluster = context.getCluster();
453 if(!cluster.isPresent()) {
457 ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
458 if(leaderActor == null) {
462 Address leaderAddress = leaderActor.anchorPath().address();
464 CurrentClusterState state = cluster.get().state();
465 Set<Member> unreachable = state.getUnreachable();
467 LOG.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
470 for(Member m: unreachable) {
471 if(leaderAddress.equals(m.address())) {
472 LOG.info("{}: Leader {} is unreachable", logName(), leaderAddress);
477 for(Member m: state.getMembers()) {
478 if(leaderAddress.equals(m.address())) {
479 if(m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
480 LOG.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
481 leaderAddress, m.status());
484 LOG.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
485 leaderAddress, m.status());
491 LOG.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
496 private void handleInstallSnapshot(final ActorRef sender, InstallSnapshot installSnapshot) {
498 LOG.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
500 leaderId = installSnapshot.getLeaderId();
502 if(snapshotTracker == null){
503 snapshotTracker = new SnapshotTracker(LOG, installSnapshot.getTotalChunks());
506 updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
509 final InstallSnapshotReply reply = new InstallSnapshotReply(
510 currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
512 if(snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
513 installSnapshot.getLastChunkHashCode())){
514 Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(),
515 new ArrayList<ReplicatedLogEntry>(),
516 installSnapshot.getLastIncludedIndex(),
517 installSnapshot.getLastIncludedTerm(),
518 installSnapshot.getLastIncludedIndex(),
519 installSnapshot.getLastIncludedTerm(),
520 context.getTermInformation().getCurrentTerm(),
521 context.getTermInformation().getVotedFor(),
522 installSnapshot.getServerConfig().orNull());
524 ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
526 public void onSuccess() {
527 LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
529 sender.tell(reply, actor());
533 public void onFailure() {
534 sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor());
538 actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
540 snapshotTracker = null;
542 LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
544 sender.tell(reply, actor());
546 } catch (SnapshotTracker.InvalidChunkException e) {
547 LOG.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
549 sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
550 -1, false), actor());
551 snapshotTracker = null;
553 } catch (Exception e){
554 LOG.error("{}: Exception in InstallSnapshot of follower", logName(), e);
556 //send reply with success as false. The chunk will be sent again on failure
557 sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
558 installSnapshot.getChunkIndex(), false), actor());
564 public void close() {
569 SnapshotTracker getSnapshotTracker(){
570 return snapshotTracker;