933fde98b9b432f60b545541d0b05cc03ef8da0d
[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.ApplyJournalEntries;
38 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
39 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
40 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
41 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
42 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
43 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
44 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
45 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
46 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
47 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
48 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
49 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
50 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
51 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
52 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
53 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
54 import org.opendaylight.yangtools.concepts.Identifier;
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     /**
209      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
210      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
211      * implementation for messages which they do not handle
212      *
213      * @param message Incoming command message
214      */
215     protected void handleNonRaftCommand(final Object message) {
216         unhandled(message);
217     }
218
219     /**
220      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
221      * {@link #handleNonRaftCommand(Object)} instead.
222      */
223     @Deprecated
224     @Override
225     // FIXME: make this method final once our unit tests do not need to override it
226     protected void handleCommand(final Object message) {
227         if (serverConfigurationSupport.handleMessage(message, getSender())) {
228             return;
229         }
230         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
231             return;
232         }
233
234         if (message instanceof ApplyState) {
235             ApplyState applyState = (ApplyState) message;
236
237             long startTime = System.nanoTime();
238
239             if(LOG.isDebugEnabled()) {
240                 LOG.debug("{}: Applying state for log index {} data {}",
241                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
242                     applyState.getReplicatedLogEntry().getData());
243             }
244
245             applyState(applyState.getClientActor(), applyState.getIdentifier(),
246                 applyState.getReplicatedLogEntry().getData());
247
248             long elapsedTime = System.nanoTime() - startTime;
249             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
250                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
251                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
252             }
253
254             if (!hasFollowers()) {
255                 // for single node, the capture should happen after the apply state
256                 // as we delete messages from the persistent journal which have made it to the snapshot
257                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
258                 // and recovery shows data missing
259                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
260
261                 context.getSnapshotManager().trimLog(context.getLastApplied());
262             }
263
264         } else if (message instanceof ApplyJournalEntries) {
265             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
266             if(LOG.isDebugEnabled()) {
267                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
268             }
269
270             persistence().persist(applyEntries, NoopProcedure.instance());
271
272         } else if (message instanceof FindLeader) {
273             getSender().tell(
274                 new FindLeaderReply(getLeaderAddress()),
275                 getSelf()
276             );
277         } else if(message instanceof GetOnDemandRaftState) {
278             onGetOnDemandRaftStats();
279         } else if(message instanceof InitiateCaptureSnapshot) {
280             captureSnapshot();
281         } else if(message instanceof SwitchBehavior) {
282             switchBehavior(((SwitchBehavior) message));
283         } else if(message instanceof LeaderTransitioning) {
284             onLeaderTransitioning();
285         } else if(message instanceof Shutdown) {
286             onShutDown();
287         } else if(message instanceof Runnable) {
288             ((Runnable)message).run();
289         } else {
290             // Processing the message may affect the state, hence we need to capture it
291             final RaftActorBehavior currentBehavior = getCurrentBehavior();
292             final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
293
294             // A behavior indicates that it processed the change by returning a reference to the next behavior
295             // to be used. A null return indicates it has not processed the message and we should be passing it to
296             // the subclass for handling.
297             final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
298             if (nextBehavior != null) {
299                 switchBehavior(state, nextBehavior);
300             } else {
301                 handleNonRaftCommand(message);
302             }
303         }
304     }
305
306     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
307         LOG.debug("{}: Initiating leader transfer", persistenceId());
308
309         if(leadershipTransferInProgress == null) {
310             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
311             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
312                 @Override
313                 public void onSuccess(ActorRef raftActorRef) {
314                     leadershipTransferInProgress = null;
315                 }
316
317                 @Override
318                 public void onFailure(ActorRef raftActorRef) {
319                     leadershipTransferInProgress = null;
320                 }
321             });
322
323             leadershipTransferInProgress.addOnComplete(onComplete);
324             leadershipTransferInProgress.init();
325         } else {
326             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
327             leadershipTransferInProgress.addOnComplete(onComplete);
328         }
329     }
330
331     private void onShutDown() {
332         LOG.debug("{}: onShutDown", persistenceId());
333
334         if(shuttingDown) {
335             return;
336         }
337
338         shuttingDown = true;
339
340         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
341         if (currentBehavior.state() != RaftState.Leader) {
342             // For non-leaders shutdown is a no-op
343             self().tell(PoisonPill.getInstance(), self());
344             return;
345         }
346
347         if (context.hasFollowers()) {
348             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
349                 @Override
350                 public void onSuccess(ActorRef raftActorRef) {
351                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
352                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
353                 }
354
355                 @Override
356                 public void onFailure(ActorRef raftActorRef) {
357                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
358                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
359                 }
360             });
361         } else {
362             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
363                 @Override
364                 protected void doRun() {
365                     self().tell(PoisonPill.getInstance(), self());
366                 }
367
368                 @Override
369                 protected void doCancel() {
370                     self().tell(PoisonPill.getInstance(), self());
371                 }
372             });
373         }
374     }
375
376     private void onLeaderTransitioning() {
377         LOG.debug("{}: onLeaderTransitioning", persistenceId());
378         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
379         if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
380             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
381                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
382         }
383     }
384
385     private void switchBehavior(SwitchBehavior message) {
386         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
387             RaftState newState = message.getNewState();
388             if( newState == RaftState.Leader || newState == RaftState.Follower) {
389                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
390                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
391                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
392             } else {
393                 LOG.warn("Switching to behavior : {} - not supported", newState);
394             }
395         }
396     }
397
398     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
399         setCurrentBehavior(nextBehavior);
400         handleBehaviorChange(oldBehaviorState, nextBehavior);
401     }
402
403     @VisibleForTesting
404     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
405         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
406     }
407
408     private void onGetOnDemandRaftStats() {
409         // Debugging message to retrieve raft stats.
410
411         Map<String, String> peerAddresses = new HashMap<>();
412         Map<String, Boolean> peerVotingStates = new HashMap<>();
413         for(PeerInfo info: context.getPeers()) {
414             peerVotingStates.put(info.getId(), info.isVoting());
415             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
416         }
417
418         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
419         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
420                 .commitIndex(context.getCommitIndex())
421                 .currentTerm(context.getTermInformation().getCurrentTerm())
422                 .inMemoryJournalDataSize(replicatedLog().dataSize())
423                 .inMemoryJournalLogSize(replicatedLog().size())
424                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
425                 .lastApplied(context.getLastApplied())
426                 .lastIndex(replicatedLog().lastIndex())
427                 .lastTerm(replicatedLog().lastTerm())
428                 .leader(getLeaderId())
429                 .raftState(currentBehavior.state().toString())
430                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
431                 .snapshotIndex(replicatedLog().getSnapshotIndex())
432                 .snapshotTerm(replicatedLog().getSnapshotTerm())
433                 .votedFor(context.getTermInformation().getVotedFor())
434                 .isVoting(context.isVotingMember())
435                 .peerAddresses(peerAddresses)
436                 .peerVotingStates(peerVotingStates)
437                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
438
439         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
440         if (lastLogEntry != null) {
441             builder.lastLogIndex(lastLogEntry.getIndex());
442             builder.lastLogTerm(lastLogEntry.getTerm());
443         }
444
445         if(getCurrentBehavior() instanceof AbstractLeader) {
446             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
447             Collection<String> followerIds = leader.getFollowerIds();
448             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
449             for(String id: followerIds) {
450                 final FollowerLogInformation info = leader.getFollower(id);
451                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
452                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
453                         context.getPeerInfo(info.getId()).isVoting()));
454             }
455
456             builder.followerInfoList(followerInfoList);
457         }
458
459         sender().tell(builder.build(), self());
460
461     }
462
463     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
464         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
465
466         if (oldBehavior != currentBehavior){
467             onStateChanged();
468         }
469
470         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
471         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
472
473         // it can happen that the state has not changed but the leader has changed.
474         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
475         if(!Objects.equals(lastValidLeaderId, currentBehavior.getLeaderId()) ||
476            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
477             if(roleChangeNotifier.isPresent()) {
478                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
479                         currentBehavior.getLeaderPayloadVersion()), getSelf());
480             }
481
482             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
483
484             if(leadershipTransferInProgress != null) {
485                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
486             }
487
488             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
489         }
490
491         if (roleChangeNotifier.isPresent() &&
492                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
493             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
494                     currentBehavior.state().name()), getSelf());
495         }
496     }
497
498     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
499         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
500     }
501
502     @Override
503     public long snapshotSequenceNr() {
504         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
505         // so that we can delete the persistent journal based on the saved sequence-number
506         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
507         // was saved and not the number we saved.
508         // We would want to override it , by asking akka to use the last-sequence number known to us.
509         return context.getSnapshotManager().getLastSequenceNumber();
510     }
511
512     /**
513      * When a derived RaftActor needs to persist something it must call
514      * persistData.
515      *
516      * @param clientActor
517      * @param identifier
518      * @param data
519      */
520     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
521
522         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
523             context.getReplicatedLog().lastIndex() + 1,
524             context.getTermInformation().getCurrentTerm(), data);
525
526         if(LOG.isDebugEnabled()) {
527             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
528         }
529
530         final RaftActorContext raftContext = getRaftActorContext();
531
532         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
533             if (!hasFollowers()){
534                 // Increment the Commit Index and the Last Applied values
535                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
536                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
537
538                 // Apply the state immediately.
539                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
540
541                 // Send a ApplyJournalEntries message so that we write the fact that we applied
542                 // the state to durable storage
543                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
544
545             } else if (clientActor != null) {
546                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
547
548                 // Send message for replication
549                 getCurrentBehavior().handleMessage(getSelf(),
550                         new Replicate(clientActor, identifier, replicatedLogEntry1));
551             }
552         });
553     }
554
555     private ReplicatedLog replicatedLog() {
556         return context.getReplicatedLog();
557     }
558
559     protected String getId() {
560         return context.getId();
561     }
562
563     @VisibleForTesting
564     void setCurrentBehavior(RaftActorBehavior behavior) {
565         context.setCurrentBehavior(behavior);
566     }
567
568     protected RaftActorBehavior getCurrentBehavior() {
569         return context.getCurrentBehavior();
570     }
571
572     /**
573      * Derived actors can call the isLeader method to check if the current
574      * RaftActor is the Leader or not
575      *
576      * @return true it this RaftActor is a Leader false otherwise
577      */
578     protected boolean isLeader() {
579         return context.getId().equals(getCurrentBehavior().getLeaderId());
580     }
581
582     protected final boolean isLeaderActive() {
583         return getRaftState() != RaftState.IsolatedLeader && !shuttingDown &&
584                 !isLeadershipTransferInProgress();
585     }
586
587     private boolean isLeadershipTransferInProgress() {
588         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
589     }
590
591     /**
592      * Derived actor can call getLeader if they need a reference to the Leader.
593      * This would be useful for example in forwarding a request to an actor
594      * which is the leader
595      *
596      * @return A reference to the leader if known, null otherwise
597      */
598     protected ActorSelection getLeader(){
599         String leaderAddress = getLeaderAddress();
600
601         if(leaderAddress == null){
602             return null;
603         }
604
605         return context.actorSelection(leaderAddress);
606     }
607
608     /**
609      *
610      * @return the current leader's id
611      */
612     protected final String getLeaderId(){
613         return getCurrentBehavior().getLeaderId();
614     }
615
616     @VisibleForTesting
617     protected final RaftState getRaftState() {
618         return getCurrentBehavior().state();
619     }
620
621     protected Long getCurrentTerm(){
622         return context.getTermInformation().getCurrentTerm();
623     }
624
625     protected RaftActorContext getRaftActorContext() {
626         return context;
627     }
628
629     protected void updateConfigParams(ConfigParams configParams) {
630
631         // obtain the RaftPolicy for oldConfigParams and the updated one.
632         String oldRaftPolicy = context.getConfigParams().
633             getCustomRaftPolicyImplementationClass();
634         String newRaftPolicy = configParams.
635             getCustomRaftPolicyImplementationClass();
636
637         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
638             oldRaftPolicy, newRaftPolicy);
639         context.setConfigParams(configParams);
640         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
641             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
642             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
643             // avoids potential disruption. Otherwise, switch to Follower normally.
644             RaftActorBehavior behavior = getCurrentBehavior();
645             if (behavior != null && behavior.state() == RaftState.Follower) {
646                 String previousLeaderId = behavior.getLeaderId();
647                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
648
649                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId);
650
651                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
652             } else {
653                 initializeBehavior();
654             }
655         }
656     }
657
658     public final DataPersistenceProvider persistence() {
659         return delegatingPersistenceProvider.getDelegate();
660     }
661
662     public void setPersistence(DataPersistenceProvider provider) {
663         delegatingPersistenceProvider.setDelegate(provider);
664     }
665
666     protected void setPersistence(boolean persistent) {
667         if(persistent) {
668             setPersistence(new PersistentDataProvider(this));
669         } else {
670             setPersistence(new NonPersistentDataProvider() {
671                 /**
672                  * The way snapshotting works is,
673                  * <ol>
674                  * <li> RaftActor calls createSnapshot on the Shard
675                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
676                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
677                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
678                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
679                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
680                  * </ol>
681                  */
682                 @Override
683                 public void saveSnapshot(Object o) {
684                     // Make saving Snapshot successful
685                     // Committing the snapshot here would end up calling commit in the creating state which would
686                     // be a state violation. That's why now we send a message to commit the snapshot.
687                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
688                 }
689             });
690         }
691     }
692
693     /**
694      * setPeerAddress sets the address of a known peer at a later time.
695      * <p>
696      * This is to account for situations where a we know that a peer
697      * exists but we do not know an address up-front. This may also be used in
698      * situations where a known peer starts off in a different location and we
699      * need to change it's address
700      * <p>
701      * Note that if the peerId does not match the list of peers passed to
702      * this actor during construction an IllegalStateException will be thrown.
703      *
704      * @param peerId
705      * @param peerAddress
706      */
707     protected void setPeerAddress(String peerId, String peerAddress){
708         context.setPeerAddress(peerId, peerAddress);
709     }
710
711     /**
712      * The applyState method will be called by the RaftActor when some data
713      * needs to be applied to the actor's state
714      *
715      * @param clientActor A reference to the client who sent this message. This
716      *                    is the same reference that was passed to persistData
717      *                    by the derived actor. clientActor may be null when
718      *                    the RaftActor is behaving as a follower or during
719      *                    recovery.
720      * @param identifier  The identifier of the persisted data. This is also
721      *                    the same identifier that was passed to persistData by
722      *                    the derived actor. identifier may be null when
723      *                    the RaftActor is behaving as a follower or during
724      *                    recovery
725      * @param data        A piece of data that was persisted by the persistData call.
726      *                    This should NEVER be null.
727      */
728     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
729
730     /**
731      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
732      */
733     @Nonnull
734     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
735
736     /**
737      * This method is called when recovery is complete.
738      */
739     protected abstract void onRecoveryComplete();
740
741     /**
742      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
743      */
744     @Nonnull
745     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
746
747     /**
748      * This method will be called by the RaftActor when the state of the
749      * RaftActor changes. The derived actor can then use methods like
750      * isLeader or getLeader to do something useful
751      */
752     protected abstract void onStateChanged();
753
754     /**
755      * Notifier Actor for this RaftActor to notify when a role change happens
756      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
757      */
758     protected abstract Optional<ActorRef> getRoleChangeNotifier();
759
760     /**
761      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
762      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
763      * work prior to performing the operation. On completion of any work, the run method must be called on the
764      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
765      * this actor's thread dispatcher as as it modifies internal state.
766      * <p>
767      * The default implementation immediately runs the operation.
768      *
769      * @param operation the operation to run
770      */
771     protected void pauseLeader(Runnable operation) {
772         operation.run();
773     }
774
775     protected void onLeaderChanged(String oldLeader, String newLeader) {
776
777     };
778
779     private String getLeaderAddress(){
780         if(isLeader()){
781             return getSelf().path().toString();
782         }
783         String leaderId = getLeaderId();
784         if (leaderId == null) {
785             return null;
786         }
787         String peerAddress = context.getPeerAddress(leaderId);
788         if(LOG.isDebugEnabled()) {
789             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
790                     persistenceId(), leaderId, peerAddress);
791         }
792
793         return peerAddress;
794     }
795
796     protected boolean hasFollowers(){
797         return getRaftActorContext().hasFollowers();
798     }
799
800     private void captureSnapshot() {
801         SnapshotManager snapshotManager = context.getSnapshotManager();
802
803         if (!snapshotManager.isCapturing()) {
804             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
805             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
806                 replicatedLog().last(), idx);
807
808             snapshotManager.capture(replicatedLog().last(), idx);
809         }
810     }
811
812     /**
813      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
814      * in which case we need to step down.
815      */
816     void becomeNonVoting() {
817         if (isLeader()) {
818             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
819                 @Override
820                 public void onSuccess(ActorRef raftActorRef) {
821                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
822                     ensureFollowerState();
823                 }
824
825                 @Override
826                 public void onFailure(ActorRef raftActorRef) {
827                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
828                     ensureFollowerState();
829                 }
830
831                 private void ensureFollowerState() {
832                     // Whether or not leadership transfer succeeded, we have to step down as leader and
833                     // switch to Follower so ensure that.
834                     if (getRaftState() != RaftState.Follower) {
835                         initializeBehavior();
836                     }
837                 }
838             });
839         }
840     }
841
842     /**
843      * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
844      *             whose type for fromIndex is long instead of int. This class was kept for backwards
845      *             compatibility with Helium.
846      */
847     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
848     @SuppressWarnings("serial")
849     @Deprecated
850     static class DeleteEntries implements Serializable {
851         private final int fromIndex;
852
853         public DeleteEntries(int fromIndex) {
854             this.fromIndex = fromIndex;
855         }
856
857         public int getFromIndex() {
858             return fromIndex;
859         }
860     }
861
862     /**
863      * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
864      *             which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
865      */
866     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
867     @SuppressWarnings("serial")
868     @Deprecated
869     static class UpdateElectionTerm implements Serializable {
870         private final long currentTerm;
871         private final String votedFor;
872
873         public UpdateElectionTerm(long currentTerm, String votedFor) {
874             this.currentTerm = currentTerm;
875             this.votedFor = votedFor;
876         }
877
878         public long getCurrentTerm() {
879             return currentTerm;
880         }
881
882         public String getVotedFor() {
883             return votedFor;
884         }
885     }
886
887     /**
888      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
889      */
890     private static abstract class BehaviorState implements Immutable {
891         @Nullable abstract RaftActorBehavior getBehavior();
892         @Nullable abstract String getLastValidLeaderId();
893         @Nullable abstract short getLeaderPayloadVersion();
894     }
895
896     /**
897      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
898      */
899     private static final class SimpleBehaviorState extends BehaviorState {
900         private final RaftActorBehavior behavior;
901         private final String lastValidLeaderId;
902         private final short leaderPayloadVersion;
903
904         SimpleBehaviorState(final String lastValidLeaderId, final RaftActorBehavior behavior) {
905             this.lastValidLeaderId = lastValidLeaderId;
906             this.behavior = Preconditions.checkNotNull(behavior);
907             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
908         }
909
910         @Override
911         RaftActorBehavior getBehavior() {
912             return behavior;
913         }
914
915         @Override
916         String getLastValidLeaderId() {
917             return lastValidLeaderId;
918         }
919
920         @Override
921         short getLeaderPayloadVersion() {
922             return leaderPayloadVersion;
923         }
924     }
925
926     /**
927      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
928      * An instance is created for each RaftActor. It has two functions:
929      * - it keeps track of the last leader ID we have encountered since we have been created
930      * - it creates state capture needed to transition from one behavior to the next
931      */
932     private static final class BehaviorStateTracker {
933         /**
934          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
935          * allowed before we receive the first message, we know the leader ID to be null.
936          */
937         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
938             @Override
939             RaftActorBehavior getBehavior() {
940                 return null;
941             }
942
943             @Override
944             String getLastValidLeaderId() {
945                 return null;
946             }
947
948             @Override
949             short getLeaderPayloadVersion() {
950                 return -1;
951             }
952         };
953
954         private String lastValidLeaderId;
955
956         BehaviorState capture(final RaftActorBehavior behavior) {
957             if (behavior == null) {
958                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
959                 return NULL_BEHAVIOR_STATE;
960             }
961
962             final String leaderId = behavior.getLeaderId();
963             if (leaderId != null) {
964                 lastValidLeaderId = leaderId;
965             }
966
967             return new SimpleBehaviorState(lastValidLeaderId, behavior);
968         }
969     }
970
971 }