7e8a7725910d599f7cd56ef40daaa06de2f8bece
[controller.git] / 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 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;
23 import java.util.Set;
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;
41
42 /**
43  * The behavior of a RaftActor in the Follower raft state.
44  * <ul>
45  * <li> Respond to RPCs from candidates and leaders
46  * <li> If election timeout elapses without receiving AppendEntries
47  * RPC from current leader or granting vote to candidate:
48  * convert to candidate
49  * </ul>
50  */
51 public class Follower extends AbstractRaftActorBehavior {
52     private static final int SYNC_THRESHOLD = 10;
53
54     private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
55
56     private final SyncStatusTracker initialSyncStatusTracker;
57
58     private final Procedure<ReplicatedLogEntry> appendAndPersistCallback =
59         logEntry -> context.getReplicatedLog().captureSnapshotIfReady(logEntry);
60
61     private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted();
62     private SnapshotTracker snapshotTracker = null;
63     private String leaderId;
64     private short leaderPayloadVersion;
65
66     public Follower(RaftActorContext context) {
67         this(context, null, (short)-1);
68     }
69
70     public Follower(RaftActorContext context, String initialLeaderId, short initialLeaderPayloadVersion) {
71         super(context, RaftState.Follower);
72         this.leaderId = initialLeaderId;
73         this.leaderPayloadVersion = initialLeaderPayloadVersion;
74
75         initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), SYNC_THRESHOLD);
76
77         if (context.getPeerIds().isEmpty() && getLeaderId() == null) {
78             actor().tell(TimeoutNow.INSTANCE, actor());
79         } else {
80             scheduleElection(electionDuration());
81         }
82     }
83
84     @Override
85     public final String getLeaderId() {
86         return leaderId;
87     }
88
89     @VisibleForTesting
90     protected final void setLeaderId(@Nullable final String leaderId) {
91         this.leaderId = leaderId;
92     }
93
94     @Override
95     public short getLeaderPayloadVersion() {
96         return leaderPayloadVersion;
97     }
98
99     @VisibleForTesting
100     protected final void setLeaderPayloadVersion(short leaderPayloadVersion) {
101         this.leaderPayloadVersion = leaderPayloadVersion;
102     }
103
104     private void restartLastLeaderMessageTimer() {
105         if (lastLeaderMessageTimer.isRunning()) {
106             lastLeaderMessageTimer.reset();
107         }
108
109         lastLeaderMessageTimer.start();
110     }
111
112     private boolean isLogEntryPresent(long index) {
113         if (context.getReplicatedLog().isInSnapshot(index)) {
114             return true;
115         }
116
117         ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
118         return entry != null;
119
120     }
121
122     private void updateInitialSyncStatus(long currentLeaderCommit, String newLeaderId) {
123         initialSyncStatusTracker.update(newLeaderId, currentLeaderCommit, context.getCommitIndex());
124     }
125
126     @Override
127     protected RaftActorBehavior handleAppendEntries(ActorRef sender, AppendEntries appendEntries) {
128
129         int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
130         if (log.isTraceEnabled()) {
131             log.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
132         } else if (log.isDebugEnabled() && numLogEntries > 0) {
133             log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
134         }
135
136         // TODO : Refactor this method into a bunch of smaller methods
137         // to make it easier to read. Before refactoring ensure tests
138         // cover the code properly
139
140         if (snapshotTracker != null && !snapshotTracker.getLeaderId().equals(appendEntries.getLeaderId())) {
141             log.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the "
142                 + "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId());
143             snapshotTracker = null;
144         }
145
146         if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
147             // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
148             AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
149                     lastIndex(), lastTerm(), context.getPayloadVersion());
150
151             log.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
152             sender.tell(reply, actor());
153
154             return this;
155         }
156
157         // If we got here then we do appear to be talking to the leader
158         leaderId = appendEntries.getLeaderId();
159         leaderPayloadVersion = appendEntries.getPayloadVersion();
160
161         updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
162         // First check if the logs are in sync or not
163         long lastIndex = lastIndex();
164
165         if (isOutOfSync(appendEntries)) {
166             // We found that the log was out of sync so just send a negative
167             // reply and return
168
169             log.debug("{}: Follower is out-of-sync, so sending negative reply, lastIndex: {}, lastTerm: {}",
170                         logName(), lastIndex, lastTerm());
171
172             sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
173                     lastTerm(), context.getPayloadVersion()), actor());
174             return this;
175         }
176
177         if (appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
178
179             log.debug("{}: Number of entries to be appended = {}", logName(),
180                         appendEntries.getEntries().size());
181
182             // 3. If an existing entry conflicts with a new one (same index
183             // but different terms), delete the existing entry and all that
184             // follow it (§5.3)
185             int addEntriesFrom = 0;
186             if (context.getReplicatedLog().size() > 0) {
187
188                 // Find the entry up until the one that is not in the follower's log
189                 for (int i = 0;i < appendEntries.getEntries().size(); i++, addEntriesFrom++) {
190                     ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
191
192                     if (!isLogEntryPresent(matchEntry.getIndex())) {
193                         // newEntry not found in the log
194                         break;
195                     }
196
197                     long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex());
198
199                     log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry,
200                             existingEntryTerm);
201
202                     // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know
203                     // what the term was so we'll assume it matches.
204                     if (existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) {
205                         continue;
206                     }
207
208                     if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
209
210                         log.debug("{}: Removing entries from log starting at {}", logName(),
211                                 matchEntry.getIndex());
212
213                         // Entries do not match so remove all subsequent entries
214                         if (!context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) {
215                             // Could not remove the entries - this means the matchEntry index must be in the
216                             // snapshot and not the log. In this case the prior entries are part of the state
217                             // so we must send back a reply to force a snapshot to completely re-sync the
218                             // follower's log and state.
219
220                             log.debug("{}: Could not remove entries - sending reply to force snapshot", logName());
221                             sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
222                                     lastTerm(), context.getPayloadVersion(), true), actor());
223                             return this;
224                         }
225
226                         break;
227                     } else {
228                         sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
229                                 lastTerm(), context.getPayloadVersion(), true), actor());
230                         return this;
231                     }
232                 }
233             }
234
235             lastIndex = lastIndex();
236             log.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(),
237                     lastIndex, addEntriesFrom);
238
239             // 4. Append any new entries not already in the log
240             for (int i = addEntriesFrom; i < appendEntries.getEntries().size(); i++) {
241                 ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
242
243                 log.debug("{}: Append entry to log {}", logName(), entry.getData());
244
245                 context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback, false);
246
247                 if (entry.getData() instanceof ServerConfigurationPayload) {
248                     context.updatePeerIds((ServerConfigurationPayload)entry.getData());
249                 }
250             }
251
252             log.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
253         }
254
255         // 5. If leaderCommit > commitIndex, set commitIndex =
256         // min(leaderCommit, index of last new entry)
257
258         lastIndex = lastIndex();
259         long prevCommitIndex = context.getCommitIndex();
260
261         if (appendEntries.getLeaderCommit() > prevCommitIndex) {
262             context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
263         }
264
265         if (prevCommitIndex != context.getCommitIndex()) {
266             log.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
267         }
268
269         // If commitIndex > lastApplied: increment lastApplied, apply
270         // log[lastApplied] to state machine (§5.3)
271         // check if there are any entries to be applied. last-applied can be equal to last-index
272         if (appendEntries.getLeaderCommit() > context.getLastApplied()
273                 && context.getLastApplied() < lastIndex) {
274             if (log.isDebugEnabled()) {
275                 log.debug("{}: applyLogToStateMachine, appendEntries.getLeaderCommit(): {},"
276                         + "context.getLastApplied(): {}, lastIndex(): {}", logName(),
277                     appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
278             }
279
280             applyLogToStateMachine(appendEntries.getLeaderCommit());
281         }
282
283         AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
284             lastIndex, lastTerm(), context.getPayloadVersion());
285
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);
290         }
291
292         sender.tell(reply, actor());
293
294         if (!context.getSnapshotManager().isCapturing()) {
295             super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
296         }
297
298         return this;
299     }
300
301     private boolean isOutOfSync(AppendEntries appendEntries) {
302
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;
308
309         if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
310
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
313             // it's log.
314
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) {
318
319             // The follower's log is out of sync because the Leader's
320             // prevLogIndex entry was not found in it's log
321
322             log.debug("{}: The log is not empty but the prevLogIndex {} was not found in it - "
323                     + "lastIndex: {}, snapshotIndex: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex,
324                     context.getReplicatedLog().getSnapshotIndex());
325         } else if (lastIndex > -1 && prevEntryPresent && prevLogTerm != appendEntries.getPrevLogTerm()) {
326
327             // The follower's log is out of sync because the Leader's
328             // prevLogIndex entry does exist in the follower's log but it has
329             // a different term in it
330
331             log.debug("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append entries"
332                       + "prevLogTerm {} - lastIndex: {}, snapshotIndex: {}", logName(), appendEntries.getPrevLogIndex(),
333                       prevLogTerm, appendEntries.getPrevLogTerm(), lastIndex,
334                       context.getReplicatedLog().getSnapshotIndex());
335         } else if (appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
336                 && appendEntries.getReplicatedToAllIndex() != -1
337                 && !isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
338             // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
339             // the previous entry in it's in-memory journal
340
341             log.debug("{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the"
342                     + " in-memory journal", logName(), appendEntries.getReplicatedToAllIndex());
343         } else if (appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
344                 && appendEntries.getReplicatedToAllIndex() != -1 && numLogEntries > 0
345                 && !isLogEntryPresent(appendEntries.getEntries().get(0).getIndex() - 1)) {
346             log.debug("{}: Cannot append entries because the calculated previousIndex {} was not found in the "
347                     + " in-memory journal", logName(), appendEntries.getEntries().get(0).getIndex() - 1);
348         } else {
349             outOfSync = false;
350         }
351         return outOfSync;
352     }
353
354     @Override
355     protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
356         AppendEntriesReply appendEntriesReply) {
357         return this;
358     }
359
360     @Override
361     protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
362         RequestVoteReply requestVoteReply) {
363         return this;
364     }
365
366     @Override
367     public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
368         if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
369             return handleElectionTimeout(message);
370         }
371
372         if (!(message instanceof RaftRPC)) {
373             // The rest of the processing requires the message to be a RaftRPC
374             return null;
375         }
376
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());
384
385             context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
386         }
387
388         if (rpc instanceof InstallSnapshot) {
389             handleInstallSnapshot(sender, (InstallSnapshot) rpc);
390             restartLastLeaderMessageTimer();
391             scheduleElection(electionDuration());
392             return this;
393         }
394
395         if (!(rpc instanceof RequestVote) || canGrantVote((RequestVote) rpc)) {
396             restartLastLeaderMessageTimer();
397             scheduleElection(electionDuration());
398         }
399
400         return super.handleMessage(sender, rpc);
401     }
402
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;
413
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
422                 // to Candidate,
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());
427                 } else {
428                     log.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
429                     return internalSwitchBehavior(RaftState.Candidate);
430                 }
431             } else {
432                 log.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
433                         logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
434                 scheduleElection(electionDuration());
435             }
436         } else if (message instanceof ElectionTimeout) {
437             if (noLeaderMessageReceived) {
438                 setLeaderId(null);
439             }
440
441             scheduleElection(electionDuration());
442         }
443
444         return this;
445     }
446
447     private boolean isLeaderAvailabilityKnown() {
448         if (leaderId == null) {
449             return false;
450         }
451
452         Optional<Cluster> cluster = context.getCluster();
453         if (!cluster.isPresent()) {
454             return false;
455         }
456
457         ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
458         if (leaderActor == null) {
459             return false;
460         }
461
462         Address leaderAddress = leaderActor.anchorPath().address();
463
464         CurrentClusterState state = cluster.get().state();
465         Set<Member> unreachable = state.getUnreachable();
466
467         log.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
468                 unreachable);
469
470         for (Member m: unreachable) {
471             if (leaderAddress.equals(m.address())) {
472                 log.info("{}: Leader {} is unreachable", logName(), leaderAddress);
473                 return false;
474             }
475         }
476
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());
482                     return true;
483                 } else {
484                     log.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
485                             leaderAddress, m.status());
486                     return false;
487                 }
488             }
489         }
490
491         log.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
492
493         return false;
494     }
495
496     private void handleInstallSnapshot(final ActorRef sender, InstallSnapshot installSnapshot) {
497
498         log.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
499
500         leaderId = installSnapshot.getLeaderId();
501
502         if (snapshotTracker == null) {
503             snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId());
504         }
505
506         updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
507
508         try {
509             final InstallSnapshotReply reply = new InstallSnapshotReply(
510                     currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
511
512             if (snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
513                     installSnapshot.getLastChunkHashCode())) {
514                 Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(),
515                         new ArrayList<>(),
516                         installSnapshot.getLastIncludedIndex(),
517                         installSnapshot.getLastIncludedTerm(),
518                         installSnapshot.getLastIncludedIndex(),
519                         installSnapshot.getLastIncludedTerm(),
520                         context.getTermInformation().getCurrentTerm(),
521                         context.getTermInformation().getVotedFor(),
522                         installSnapshot.getServerConfig().orNull());
523
524                 ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
525                     @Override
526                     public void onSuccess() {
527                         log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
528
529                         sender.tell(reply, actor());
530                     }
531
532                     @Override
533                     public void onFailure() {
534                         sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor());
535                     }
536                 };
537
538                 actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
539
540                 snapshotTracker = null;
541             } else {
542                 log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
543
544                 sender.tell(reply, actor());
545             }
546         } catch (SnapshotTracker.InvalidChunkException e) {
547             log.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
548
549             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
550                     -1, false), actor());
551             snapshotTracker = null;
552
553         }
554     }
555
556     @Override
557     public void close() {
558         stopElection();
559     }
560
561     @VisibleForTesting
562     SnapshotTracker getSnapshotTracker() {
563         return snapshotTracker;
564     }
565 }