Bug 7391: Fix out-of-order LeaderStateChange events
[controller.git] / opendaylight / md-sal / sal-akka-raft / src / main / java / org / opendaylight / controller / cluster / raft / RaftActor.java
1 /*
2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
3  * Copyright (c) 2015 Brocade Communications Systems, Inc. and others.  All rights reserved.
4  *
5  * This program and the accompanying materials are made available under the
6  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
7  * and is available at http://www.eclipse.org/legal/epl-v10.html
8  */
9
10 package org.opendaylight.controller.cluster.raft;
11
12 import akka.actor.ActorRef;
13 import akka.actor.ActorSelection;
14 import akka.actor.PoisonPill;
15 import com.google.common.annotations.VisibleForTesting;
16 import com.google.common.base.Optional;
17 import com.google.common.base.Preconditions;
18 import com.google.common.base.Verify;
19 import com.google.common.collect.Lists;
20 import java.io.Serializable;
21 import java.util.Collection;
22 import java.util.HashMap;
23 import java.util.List;
24 import java.util.Map;
25 import java.util.Objects;
26 import java.util.concurrent.TimeUnit;
27 import javax.annotation.Nonnull;
28 import javax.annotation.Nullable;
29 import org.apache.commons.lang3.time.DurationFormatUtils;
30 import org.opendaylight.controller.cluster.DataPersistenceProvider;
31 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
32 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
33 import org.opendaylight.controller.cluster.PersistentDataProvider;
34 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
35 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
36 import org.opendaylight.controller.cluster.notifications.RoleChanged;
37 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
38 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
39 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
40 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
41 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
44 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
45 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
48 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
49 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
50 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
51 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
52 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
53 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
54 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
55 import org.opendaylight.yangtools.concepts.Identifier;
56 import org.opendaylight.yangtools.concepts.Immutable;
57 import org.slf4j.Logger;
58 import org.slf4j.LoggerFactory;
59
60 /**
61  * RaftActor encapsulates a state machine that needs to be kept synchronized
62  * in a cluster. It implements the RAFT algorithm as described in the paper
63  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
64  * In Search of an Understandable Consensus Algorithm</a>
65  * <p/>
66  * RaftActor has 3 states and each state has a certain behavior associated
67  * with it. A Raft actor can behave as,
68  * <ul>
69  * <li> A Leader </li>
70  * <li> A Follower (or) </li>
71  * <li> A Candidate </li>
72  * </ul>
73  * <p/>
74  * <p/>
75  * A RaftActor MUST be a Leader in order to accept requests from clients to
76  * change the state of it's encapsulated state machine. Once a RaftActor becomes
77  * a Leader it is also responsible for ensuring that all followers ultimately
78  * have the same log and therefore the same state machine as itself.
79  * <p/>
80  * <p/>
81  * The current behavior of a RaftActor determines how election for leadership
82  * is initiated and how peer RaftActors react to request for votes.
83  * <p/>
84  * <p/>
85  * Each RaftActor also needs to know the current election term. It uses this
86  * information for a couple of things. One is to simply figure out who it
87  * voted for in the last election. Another is to figure out if the message
88  * it received to update it's state is stale.
89  * <p/>
90  * <p/>
91  * The RaftActor uses akka-persistence to store it's replicated log.
92  * Furthermore through it's behaviors a Raft Actor determines
93  * <p/>
94  * <ul>
95  * <li> when a log entry should be persisted </li>
96  * <li> when a log entry should be applied to the state machine (and) </li>
97  * <li> when a snapshot should be saved </li>
98  * </ul>
99  */
100 public abstract class RaftActor extends AbstractUntypedPersistentActor {
101
102     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
103
104     protected final Logger LOG = LoggerFactory.getLogger(getClass());
105
106     /**
107      * This context should NOT be passed directly to any other actor it is
108      * only to be consumed by the RaftActorBehaviors
109      */
110     private final RaftActorContextImpl context;
111
112     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
113
114     private final PersistentDataProvider persistentProvider;
115
116     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
117
118     private RaftActorRecoverySupport raftRecovery;
119
120     private RaftActorSnapshotMessageSupport snapshotSupport;
121
122     private RaftActorServerConfigurationSupport serverConfigurationSupport;
123
124     private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
125
126     private boolean shuttingDown;
127
128     public RaftActor(String id, Map<String, String> peerAddresses,
129          Optional<ConfigParams> configParams, short payloadVersion) {
130
131         persistentProvider = new PersistentDataProvider(this);
132         delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
133
134         context = new RaftActorContextImpl(this.getSelf(),
135             this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
136             -1, -1, peerAddresses,
137             configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl(),
138             delegatingPersistenceProvider, LOG);
139
140         context.setPayloadVersion(payloadVersion);
141         context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
142     }
143
144     @Override
145     public void preStart() throws Exception {
146         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
147                 context.getConfigParams().getJournalRecoveryLogBatchSize());
148
149         super.preStart();
150
151         snapshotSupport = newRaftActorSnapshotMessageSupport();
152         serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
153     }
154
155     @Override
156     public void postStop() {
157         context.close();
158         super.postStop();
159     }
160
161     @Override
162     protected void handleRecover(Object message) {
163         if(raftRecovery == null) {
164             raftRecovery = newRaftActorRecoverySupport();
165         }
166
167         boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
168         if(recoveryComplete) {
169             onRecoveryComplete();
170
171             initializeBehavior();
172
173             raftRecovery = null;
174         }
175     }
176
177     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
178         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
179     }
180
181     @VisibleForTesting
182     void initializeBehavior(){
183         changeCurrentBehavior(new Follower(context));
184     }
185
186     @VisibleForTesting
187     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
188         final RaftActorBehavior currentBehavior = getCurrentBehavior();
189         if (currentBehavior != null) {
190             try {
191                 currentBehavior.close();
192             } catch (Exception e) {
193                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
194             }
195         }
196
197         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
198         setCurrentBehavior(newBehavior);
199         handleBehaviorChange(state, newBehavior);
200     }
201
202     /**
203      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
204      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
205      * implementation for messages which they do not handle
206      *
207      * @param message Incoming command message
208      */
209     protected void handleNonRaftCommand(final Object message) {
210         unhandled(message);
211     }
212
213     /**
214      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
215      * {@link #handleNonRaftCommand(Object)} instead.
216      */
217     @Deprecated
218     @Override
219     // FIXME: make this method final once our unit tests do not need to override it
220     protected void handleCommand(final Object message) {
221         if (serverConfigurationSupport.handleMessage(message, getSender())) {
222             return;
223         }
224         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
225             return;
226         }
227
228         if (message instanceof ApplyState) {
229             ApplyState applyState = (ApplyState) message;
230
231             long startTime = System.nanoTime();
232
233             if(LOG.isDebugEnabled()) {
234                 LOG.debug("{}: Applying state for log index {} data {}",
235                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
236                     applyState.getReplicatedLogEntry().getData());
237             }
238
239             if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
240                 applyState(applyState.getClientActor(), applyState.getIdentifier(),
241                     applyState.getReplicatedLogEntry().getData());
242             }
243
244             long elapsedTime = System.nanoTime() - startTime;
245             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
246                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
247                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
248             }
249
250             if (!hasFollowers()) {
251                 // for single node, the capture should happen after the apply state
252                 // as we delete messages from the persistent journal which have made it to the snapshot
253                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
254                 // and recovery shows data missing
255                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
256
257                 context.getSnapshotManager().trimLog(context.getLastApplied());
258             }
259
260             // Send it to the current behavior - some behaviors like PreLeader need to be notified of ApplyState.
261             possiblyHandleBehaviorMessage(message);
262
263         } else if (message instanceof ApplyJournalEntries) {
264             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
265             if(LOG.isDebugEnabled()) {
266                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
267             }
268
269             persistence().persist(applyEntries, NoopProcedure.instance());
270
271         } else if (message instanceof FindLeader) {
272             getSender().tell(
273                 new FindLeaderReply(getLeaderAddress()),
274                 getSelf()
275             );
276         } else if(message instanceof GetOnDemandRaftState) {
277             onGetOnDemandRaftStats();
278         } else if(message instanceof InitiateCaptureSnapshot) {
279             captureSnapshot();
280         } else if(message instanceof SwitchBehavior) {
281             switchBehavior((SwitchBehavior) message);
282         } else if(message instanceof LeaderTransitioning) {
283             onLeaderTransitioning((LeaderTransitioning)message);
284         } else if(message instanceof Shutdown) {
285             onShutDown();
286         } else if(message instanceof Runnable) {
287             ((Runnable)message).run();
288         } else if(message instanceof NoopPayload) {
289             persistData(null, null, (NoopPayload)message);
290         } else if (!possiblyHandleBehaviorMessage(message)) {
291             handleNonRaftCommand(message);
292         }
293     }
294
295     private boolean possiblyHandleBehaviorMessage(final Object message) {
296         final RaftActorBehavior currentBehavior = getCurrentBehavior();
297         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
298
299         // A behavior indicates that it processed the change by returning a reference to the next behavior
300         // to be used. A null return indicates it has not processed the message and we should be passing it to
301         // the subclass for handling.
302         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
303         if (nextBehavior != null) {
304             switchBehavior(state, nextBehavior);
305             return true;
306         }
307
308         return false;
309     }
310
311     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
312         LOG.debug("{}: Initiating leader transfer", persistenceId());
313
314         if(leadershipTransferInProgress == null) {
315             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
316             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
317                 @Override
318                 public void onSuccess(ActorRef raftActorRef) {
319                     leadershipTransferInProgress = null;
320                 }
321
322                 @Override
323                 public void onFailure(ActorRef raftActorRef) {
324                     leadershipTransferInProgress = null;
325                 }
326             });
327
328             leadershipTransferInProgress.addOnComplete(onComplete);
329             leadershipTransferInProgress.init();
330         } else {
331             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
332             leadershipTransferInProgress.addOnComplete(onComplete);
333         }
334     }
335
336     private void onShutDown() {
337         LOG.debug("{}: onShutDown", persistenceId());
338
339         if(shuttingDown) {
340             return;
341         }
342
343         shuttingDown = true;
344
345         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
346         if (currentBehavior.state() != RaftState.Leader) {
347             // For non-leaders shutdown is a no-op
348             self().tell(PoisonPill.getInstance(), self());
349             return;
350         }
351
352         if (context.hasFollowers()) {
353             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
354                 @Override
355                 public void onSuccess(ActorRef raftActorRef) {
356                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
357                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
358                 }
359
360                 @Override
361                 public void onFailure(ActorRef raftActorRef) {
362                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
363                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
364                 }
365             });
366         } else {
367             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
368                 @Override
369                 protected void doRun() {
370                     self().tell(PoisonPill.getInstance(), self());
371                 }
372
373                 @Override
374                 protected void doCancel() {
375                     self().tell(PoisonPill.getInstance(), self());
376                 }
377             });
378         }
379     }
380
381     private void onLeaderTransitioning(final LeaderTransitioning leaderTransitioning) {
382         LOG.debug("{}: onLeaderTransitioning: {}", persistenceId(), leaderTransitioning);
383         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
384         if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()
385                 && leaderTransitioning.getLeaderId().equals(getCurrentBehavior().getLeaderId())) {
386             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
387                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
388         }
389     }
390
391     private void switchBehavior(SwitchBehavior message) {
392         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
393             RaftState newState = message.getNewState();
394             if( newState == RaftState.Leader || newState == RaftState.Follower) {
395                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
396                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
397                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
398             } else {
399                 LOG.warn("Switching to behavior : {} - not supported", newState);
400             }
401         }
402     }
403
404     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
405         setCurrentBehavior(nextBehavior);
406         handleBehaviorChange(oldBehaviorState, nextBehavior);
407     }
408
409     @VisibleForTesting
410     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
411         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
412     }
413
414     private void onGetOnDemandRaftStats() {
415         // Debugging message to retrieve raft stats.
416
417         Map<String, String> peerAddresses = new HashMap<>();
418         Map<String, Boolean> peerVotingStates = new HashMap<>();
419         for(PeerInfo info: context.getPeers()) {
420             peerVotingStates.put(info.getId(), info.isVoting());
421             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
422         }
423
424         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
425         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
426                 .commitIndex(context.getCommitIndex())
427                 .currentTerm(context.getTermInformation().getCurrentTerm())
428                 .inMemoryJournalDataSize(replicatedLog().dataSize())
429                 .inMemoryJournalLogSize(replicatedLog().size())
430                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
431                 .lastApplied(context.getLastApplied())
432                 .lastIndex(replicatedLog().lastIndex())
433                 .lastTerm(replicatedLog().lastTerm())
434                 .leader(getLeaderId())
435                 .raftState(currentBehavior.state().toString())
436                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
437                 .snapshotIndex(replicatedLog().getSnapshotIndex())
438                 .snapshotTerm(replicatedLog().getSnapshotTerm())
439                 .votedFor(context.getTermInformation().getVotedFor())
440                 .isVoting(context.isVotingMember())
441                 .peerAddresses(peerAddresses)
442                 .peerVotingStates(peerVotingStates)
443                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
444
445         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
446         if (lastLogEntry != null) {
447             builder.lastLogIndex(lastLogEntry.getIndex());
448             builder.lastLogTerm(lastLogEntry.getTerm());
449         }
450
451         if(getCurrentBehavior() instanceof AbstractLeader) {
452             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
453             Collection<String> followerIds = leader.getFollowerIds();
454             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
455             for(String id: followerIds) {
456                 final FollowerLogInformation info = leader.getFollower(id);
457                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
458                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
459                         context.getPeerInfo(info.getId()).isVoting()));
460             }
461
462             builder.followerInfoList(followerInfoList);
463         }
464
465         sender().tell(builder.build(), self());
466
467     }
468
469     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
470         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
471
472         if (oldBehavior != currentBehavior){
473             onStateChanged();
474         }
475
476         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
477         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
478         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
479
480         // it can happen that the state has not changed but the leader has changed.
481         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
482         if(!Objects.equals(lastLeaderId, currentBehavior.getLeaderId()) ||
483            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
484             if(roleChangeNotifier.isPresent()) {
485                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
486                         currentBehavior.getLeaderPayloadVersion()), getSelf());
487             }
488
489             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
490
491             if(leadershipTransferInProgress != null) {
492                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
493             }
494
495             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
496         }
497
498         if (roleChangeNotifier.isPresent() &&
499                 (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
500             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
501                     currentBehavior.state().name()), getSelf());
502         }
503     }
504
505     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
506         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
507     }
508
509     @Override
510     public long snapshotSequenceNr() {
511         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
512         // so that we can delete the persistent journal based on the saved sequence-number
513         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
514         // was saved and not the number we saved.
515         // We would want to override it , by asking akka to use the last-sequence number known to us.
516         return context.getSnapshotManager().getLastSequenceNumber();
517     }
518
519     /**
520      * When a derived RaftActor needs to persist something it must call
521      * persistData.
522      *
523      * @param clientActor
524      * @param identifier
525      * @param data
526      */
527     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
528
529         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
530             context.getReplicatedLog().lastIndex() + 1,
531             context.getTermInformation().getCurrentTerm(), data);
532
533         if(LOG.isDebugEnabled()) {
534             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
535         }
536
537         final RaftActorContext raftContext = getRaftActorContext();
538
539         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
540             if (!hasFollowers()){
541                 // Increment the Commit Index and the Last Applied values
542                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
543                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
544
545                 // Apply the state immediately.
546                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
547
548                 // Send a ApplyJournalEntries message so that we write the fact that we applied
549                 // the state to durable storage
550                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
551
552             } else {
553                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
554
555                 // Send message for replication
556                 getCurrentBehavior().handleMessage(getSelf(),
557                         new Replicate(clientActor, identifier, replicatedLogEntry1));
558             }
559         });
560     }
561
562     private ReplicatedLog replicatedLog() {
563         return context.getReplicatedLog();
564     }
565
566     protected String getId() {
567         return context.getId();
568     }
569
570     @VisibleForTesting
571     void setCurrentBehavior(RaftActorBehavior behavior) {
572         context.setCurrentBehavior(behavior);
573     }
574
575     protected RaftActorBehavior getCurrentBehavior() {
576         return context.getCurrentBehavior();
577     }
578
579     /**
580      * Derived actors can call the isLeader method to check if the current
581      * RaftActor is the Leader or not
582      *
583      * @return true it this RaftActor is a Leader false otherwise
584      */
585     protected boolean isLeader() {
586         return context.getId().equals(getCurrentBehavior().getLeaderId());
587     }
588
589     protected final boolean isLeaderActive() {
590         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader &&
591                 !shuttingDown && !isLeadershipTransferInProgress();
592     }
593
594     private boolean isLeadershipTransferInProgress() {
595         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
596     }
597
598     /**
599      * Derived actor can call getLeader if they need a reference to the Leader.
600      * This would be useful for example in forwarding a request to an actor
601      * which is the leader
602      *
603      * @return A reference to the leader if known, null otherwise
604      */
605     public ActorSelection getLeader(){
606         String leaderAddress = getLeaderAddress();
607
608         if(leaderAddress == null){
609             return null;
610         }
611
612         return context.actorSelection(leaderAddress);
613     }
614
615     /**
616      *
617      * @return the current leader's id
618      */
619     protected final String getLeaderId(){
620         return getCurrentBehavior().getLeaderId();
621     }
622
623     @VisibleForTesting
624     protected final RaftState getRaftState() {
625         return getCurrentBehavior().state();
626     }
627
628     protected Long getCurrentTerm(){
629         return context.getTermInformation().getCurrentTerm();
630     }
631
632     protected RaftActorContext getRaftActorContext() {
633         return context;
634     }
635
636     protected void updateConfigParams(ConfigParams configParams) {
637
638         // obtain the RaftPolicy for oldConfigParams and the updated one.
639         String oldRaftPolicy = context.getConfigParams().
640             getCustomRaftPolicyImplementationClass();
641         String newRaftPolicy = configParams.
642             getCustomRaftPolicyImplementationClass();
643
644         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
645             oldRaftPolicy, newRaftPolicy);
646         context.setConfigParams(configParams);
647         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
648             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
649             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
650             // avoids potential disruption. Otherwise, switch to Follower normally.
651             RaftActorBehavior behavior = getCurrentBehavior();
652             if (behavior != null && behavior.state() == RaftState.Follower) {
653                 String previousLeaderId = behavior.getLeaderId();
654                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
655
656                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
657
658                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
659             } else {
660                 initializeBehavior();
661             }
662         }
663     }
664
665     public final DataPersistenceProvider persistence() {
666         return delegatingPersistenceProvider.getDelegate();
667     }
668
669     public void setPersistence(DataPersistenceProvider provider) {
670         delegatingPersistenceProvider.setDelegate(provider);
671     }
672
673     protected void setPersistence(boolean persistent) {
674         DataPersistenceProvider currentPersistence = persistence();
675         if(persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
676             setPersistence(new PersistentDataProvider(this));
677
678             if(getCurrentBehavior() != null) {
679                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
680                 captureSnapshot();
681             }
682         } else if(!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
683             setPersistence(new NonPersistentDataProvider() {
684                 /**
685                  * The way snapshotting works is,
686                  * <ol>
687                  * <li> RaftActor calls createSnapshot on the Shard
688                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
689                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
690                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
691                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
692                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
693                  * </ol>
694                  */
695                 @Override
696                 public void saveSnapshot(Object o) {
697                     // Make saving Snapshot successful
698                     // Committing the snapshot here would end up calling commit in the creating state which would
699                     // be a state violation. That's why now we send a message to commit the snapshot.
700                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
701                 }
702             });
703         }
704     }
705
706     /**
707      * setPeerAddress sets the address of a known peer at a later time.
708      * <p>
709      * This is to account for situations where a we know that a peer
710      * exists but we do not know an address up-front. This may also be used in
711      * situations where a known peer starts off in a different location and we
712      * need to change it's address
713      * <p>
714      * Note that if the peerId does not match the list of peers passed to
715      * this actor during construction an IllegalStateException will be thrown.
716      *
717      * @param peerId
718      * @param peerAddress
719      */
720     protected void setPeerAddress(String peerId, String peerAddress){
721         context.setPeerAddress(peerId, peerAddress);
722     }
723
724     /**
725      * The applyState method will be called by the RaftActor when some data
726      * needs to be applied to the actor's state
727      *
728      * @param clientActor A reference to the client who sent this message. This
729      *                    is the same reference that was passed to persistData
730      *                    by the derived actor. clientActor may be null when
731      *                    the RaftActor is behaving as a follower or during
732      *                    recovery.
733      * @param identifier  The identifier of the persisted data. This is also
734      *                    the same identifier that was passed to persistData by
735      *                    the derived actor. identifier may be null when
736      *                    the RaftActor is behaving as a follower or during
737      *                    recovery
738      * @param data        A piece of data that was persisted by the persistData call.
739      *                    This should NEVER be null.
740      */
741     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
742
743     /**
744      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
745      */
746     @Nonnull
747     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
748
749     /**
750      * This method is called when recovery is complete.
751      */
752     protected abstract void onRecoveryComplete();
753
754     /**
755      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
756      */
757     @Nonnull
758     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
759
760     /**
761      * This method will be called by the RaftActor when the state of the
762      * RaftActor changes. The derived actor can then use methods like
763      * isLeader or getLeader to do something useful
764      */
765     protected abstract void onStateChanged();
766
767     /**
768      * Notifier Actor for this RaftActor to notify when a role change happens
769      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
770      */
771     protected abstract Optional<ActorRef> getRoleChangeNotifier();
772
773     /**
774      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
775      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
776      * work prior to performing the operation. On completion of any work, the run method must be called on the
777      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
778      * this actor's thread dispatcher as as it modifies internal state.
779      * <p>
780      * The default implementation immediately runs the operation.
781      *
782      * @param operation the operation to run
783      */
784     protected void pauseLeader(Runnable operation) {
785         operation.run();
786     }
787
788     protected void onLeaderChanged(String oldLeader, String newLeader) {
789
790     };
791
792     private String getLeaderAddress(){
793         if(isLeader()){
794             return getSelf().path().toString();
795         }
796         String leaderId = getLeaderId();
797         if (leaderId == null) {
798             return null;
799         }
800         String peerAddress = context.getPeerAddress(leaderId);
801         if(LOG.isDebugEnabled()) {
802             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
803                     persistenceId(), leaderId, peerAddress);
804         }
805
806         return peerAddress;
807     }
808
809     protected boolean hasFollowers(){
810         return getRaftActorContext().hasFollowers();
811     }
812
813     private void captureSnapshot() {
814         SnapshotManager snapshotManager = context.getSnapshotManager();
815
816         if (!snapshotManager.isCapturing()) {
817             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
818             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
819                 replicatedLog().last(), idx);
820
821             snapshotManager.capture(replicatedLog().last(), idx);
822         }
823     }
824
825     /**
826      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
827      * in which case we need to step down.
828      */
829     void becomeNonVoting() {
830         if (isLeader()) {
831             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
832                 @Override
833                 public void onSuccess(ActorRef raftActorRef) {
834                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
835                     ensureFollowerState();
836                 }
837
838                 @Override
839                 public void onFailure(ActorRef raftActorRef) {
840                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
841                     ensureFollowerState();
842                 }
843
844                 private void ensureFollowerState() {
845                     // Whether or not leadership transfer succeeded, we have to step down as leader and
846                     // switch to Follower so ensure that.
847                     if (getRaftState() != RaftState.Follower) {
848                         initializeBehavior();
849                     }
850                 }
851             });
852         }
853     }
854
855     /**
856      * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.persisted.DeleteEntries}
857      *             whose type for fromIndex is long instead of int. This class was kept for backwards
858      *             compatibility with Helium.
859      */
860     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
861     @SuppressWarnings("serial")
862     @Deprecated
863     static class DeleteEntries implements Serializable {
864         private final int fromIndex;
865
866         public DeleteEntries(int fromIndex) {
867             this.fromIndex = fromIndex;
868         }
869
870         public int getFromIndex() {
871             return fromIndex;
872         }
873
874         private Object readResolve() {
875             return org.opendaylight.controller.cluster.raft.persisted.DeleteEntries.createMigrated(fromIndex);
876         }
877     }
878
879     /**
880      * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm}
881      *             which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
882      */
883     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
884     @SuppressWarnings("serial")
885     @Deprecated
886     static class UpdateElectionTerm implements Serializable {
887         private final long currentTerm;
888         private final String votedFor;
889
890         public UpdateElectionTerm(long currentTerm, String votedFor) {
891             this.currentTerm = currentTerm;
892             this.votedFor = votedFor;
893         }
894
895         public long getCurrentTerm() {
896             return currentTerm;
897         }
898
899         public String getVotedFor() {
900             return votedFor;
901         }
902
903         private Object readResolve() {
904             return org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm.createMigrated(
905                     currentTerm, votedFor);
906         }
907     }
908
909     /**
910      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
911      */
912     private static abstract class BehaviorState implements Immutable {
913         @Nullable abstract RaftActorBehavior getBehavior();
914         @Nullable abstract String getLastValidLeaderId();
915         @Nullable abstract String getLastLeaderId();
916         @Nullable abstract short getLeaderPayloadVersion();
917     }
918
919     /**
920      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
921      */
922     private static final class SimpleBehaviorState extends BehaviorState {
923         private final RaftActorBehavior behavior;
924         private final String lastValidLeaderId;
925         private final String lastLeaderId;
926         private final short leaderPayloadVersion;
927
928         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
929                 final RaftActorBehavior behavior) {
930             this.lastValidLeaderId = lastValidLeaderId;
931             this.lastLeaderId = lastLeaderId;
932             this.behavior = Preconditions.checkNotNull(behavior);
933             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
934         }
935
936         @Override
937         RaftActorBehavior getBehavior() {
938             return behavior;
939         }
940
941         @Override
942         String getLastValidLeaderId() {
943             return lastValidLeaderId;
944         }
945
946         @Override
947         short getLeaderPayloadVersion() {
948             return leaderPayloadVersion;
949         }
950
951         @Override
952         String getLastLeaderId() {
953             return lastLeaderId;
954         }
955     }
956
957     /**
958      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
959      * An instance is created for each RaftActor. It has two functions:
960      * - it keeps track of the last leader ID we have encountered since we have been created
961      * - it creates state capture needed to transition from one behavior to the next
962      */
963     private static final class BehaviorStateTracker {
964         /**
965          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
966          * allowed before we receive the first message, we know the leader ID to be null.
967          */
968         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
969             @Override
970             RaftActorBehavior getBehavior() {
971                 return null;
972             }
973
974             @Override
975             String getLastValidLeaderId() {
976                 return null;
977             }
978
979             @Override
980             short getLeaderPayloadVersion() {
981                 return -1;
982             }
983
984             @Override
985             String getLastLeaderId() {
986                 return null;
987             }
988         };
989
990         private String lastValidLeaderId;
991         private String lastLeaderId;
992
993         BehaviorState capture(final RaftActorBehavior behavior) {
994             if (behavior == null) {
995                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
996                 return NULL_BEHAVIOR_STATE;
997             }
998
999             lastLeaderId = behavior.getLeaderId();
1000             if (lastLeaderId != null) {
1001                 lastValidLeaderId = lastLeaderId;
1002             }
1003
1004             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
1005         }
1006     }
1007
1008 }