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