cb07e59c5b9a44ad1c7fe778bf74de4743b576ea
[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.util.Collection;
21 import java.util.HashMap;
22 import java.util.List;
23 import java.util.Map;
24 import java.util.Objects;
25 import java.util.concurrent.TimeUnit;
26 import javax.annotation.Nonnull;
27 import javax.annotation.Nullable;
28 import org.apache.commons.lang3.time.DurationFormatUtils;
29 import org.opendaylight.controller.cluster.DataPersistenceProvider;
30 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
31 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
32 import org.opendaylight.controller.cluster.PersistentDataProvider;
33 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
34 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
35 import org.opendaylight.controller.cluster.notifications.RoleChanged;
36 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
37 import org.opendaylight.controller.cluster.raft.base.messages.InitiateCaptureSnapshot;
38 import org.opendaylight.controller.cluster.raft.base.messages.LeaderTransitioning;
39 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
40 import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior;
41 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
42 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
43 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
44 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
45 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
46 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
47 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
48 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
49 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
50 import org.opendaylight.controller.cluster.raft.client.messages.Shutdown;
51 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
52 import org.opendaylight.controller.cluster.raft.persisted.NoopPayload;
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
57 /**
58  * RaftActor encapsulates a state machine that needs to be kept synchronized
59  * in a cluster. It implements the RAFT algorithm as described in the paper
60  * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
61  * In Search of an Understandable Consensus Algorithm</a>
62  * <p/>
63  * RaftActor has 3 states and each state has a certain behavior associated
64  * with it. A Raft actor can behave as,
65  * <ul>
66  * <li> A Leader </li>
67  * <li> A Follower (or) </li>
68  * <li> A Candidate </li>
69  * </ul>
70  * <p/>
71  * <p/>
72  * A RaftActor MUST be a Leader in order to accept requests from clients to
73  * change the state of it's encapsulated state machine. Once a RaftActor becomes
74  * a Leader it is also responsible for ensuring that all followers ultimately
75  * have the same log and therefore the same state machine as itself.
76  * <p/>
77  * <p/>
78  * The current behavior of a RaftActor determines how election for leadership
79  * is initiated and how peer RaftActors react to request for votes.
80  * <p/>
81  * <p/>
82  * Each RaftActor also needs to know the current election term. It uses this
83  * information for a couple of things. One is to simply figure out who it
84  * voted for in the last election. Another is to figure out if the message
85  * it received to update it's state is stale.
86  * <p/>
87  * <p/>
88  * The RaftActor uses akka-persistence to store it's replicated log.
89  * Furthermore through it's behaviors a Raft Actor determines
90  * <p/>
91  * <ul>
92  * <li> when a log entry should be persisted </li>
93  * <li> when a log entry should be applied to the state machine (and) </li>
94  * <li> when a snapshot should be saved </li>
95  * </ul>
96  */
97 public abstract class RaftActor extends AbstractUntypedPersistentActor {
98
99     private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
100
101     /**
102      * This context should NOT be passed directly to any other actor it is
103      * only to be consumed by the RaftActorBehaviors.
104      */
105     private final RaftActorContextImpl context;
106
107     private final DelegatingPersistentDataProvider delegatingPersistenceProvider;
108
109     private final PersistentDataProvider persistentProvider;
110
111     private final BehaviorStateTracker behaviorStateTracker = new BehaviorStateTracker();
112
113     private RaftActorRecoverySupport raftRecovery;
114
115     private RaftActorSnapshotMessageSupport snapshotSupport;
116
117     private RaftActorServerConfigurationSupport serverConfigurationSupport;
118
119     private RaftActorLeadershipTransferCohort leadershipTransferInProgress;
120
121     private boolean shuttingDown;
122
123     protected RaftActor(String id, Map<String, String> peerAddresses,
124          Optional<ConfigParams> configParams, short payloadVersion) {
125
126         persistentProvider = new PersistentDataProvider(this);
127         delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider);
128
129         context = new RaftActorContextImpl(this.getSelf(),
130             this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG),
131             -1, -1, peerAddresses,
132             configParams.isPresent() ? configParams.get() : new DefaultConfigParamsImpl(),
133             delegatingPersistenceProvider, LOG);
134
135         context.setPayloadVersion(payloadVersion);
136         context.setReplicatedLog(ReplicatedLogImpl.newInstance(context));
137     }
138
139     @Override
140     public void preStart() throws Exception {
141         LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
142                 context.getConfigParams().getJournalRecoveryLogBatchSize());
143
144         super.preStart();
145
146         snapshotSupport = newRaftActorSnapshotMessageSupport();
147         serverConfigurationSupport = new RaftActorServerConfigurationSupport(this);
148     }
149
150     @Override
151     public void postStop() {
152         context.close();
153         super.postStop();
154     }
155
156     @Override
157     protected void handleRecover(Object message) {
158         if (raftRecovery == null) {
159             raftRecovery = newRaftActorRecoverySupport();
160         }
161
162         boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider);
163         if (recoveryComplete) {
164             onRecoveryComplete();
165
166             initializeBehavior();
167
168             raftRecovery = null;
169         }
170     }
171
172     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
173         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
174     }
175
176     @VisibleForTesting
177     void initializeBehavior() {
178         changeCurrentBehavior(new Follower(context));
179     }
180
181     @VisibleForTesting
182     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
183         final RaftActorBehavior currentBehavior = getCurrentBehavior();
184         if (currentBehavior != null) {
185             try {
186                 currentBehavior.close();
187             } catch (Exception e) {
188                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
189             }
190         }
191
192         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
193         setCurrentBehavior(newBehavior);
194         handleBehaviorChange(state, newBehavior);
195     }
196
197     /**
198      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
199      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
200      * implementation for messages which they do not handle
201      *
202      * @param message Incoming command message
203      */
204     protected void handleNonRaftCommand(final Object message) {
205         unhandled(message);
206     }
207
208     /**
209      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
210      * {@link #handleNonRaftCommand(Object)} instead.
211      */
212     @Deprecated
213     @Override
214     // FIXME: make this method final once our unit tests do not need to override it
215     protected void handleCommand(final Object message) {
216         if (serverConfigurationSupport.handleMessage(message, getSender())) {
217             return;
218         }
219         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
220             return;
221         }
222
223         if (message instanceof ApplyState) {
224             ApplyState applyState = (ApplyState) message;
225
226             long startTime = System.nanoTime();
227
228             if (LOG.isDebugEnabled()) {
229                 LOG.debug("{}: Applying state for log index {} data {}",
230                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
231                     applyState.getReplicatedLogEntry().getData());
232             }
233
234             if (!(applyState.getReplicatedLogEntry().getData() instanceof NoopPayload)) {
235                 applyState(applyState.getClientActor(), applyState.getIdentifier(),
236                     applyState.getReplicatedLogEntry().getData());
237             }
238
239             long elapsedTime = System.nanoTime() - startTime;
240             if (elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS) {
241                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
242                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
243             }
244
245             if (!hasFollowers()) {
246                 // for single node, the capture should happen after the apply state
247                 // as we delete messages from the persistent journal which have made it to the snapshot
248                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
249                 // and recovery shows data missing
250                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
251
252                 context.getSnapshotManager().trimLog(context.getLastApplied());
253             }
254
255             // Send it to the current behavior - some behaviors like PreLeader need to be notified of ApplyState.
256             possiblyHandleBehaviorMessage(message);
257
258         } else if (message instanceof ApplyJournalEntries) {
259             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
260             LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
261
262             persistence().persist(applyEntries, NoopProcedure.instance());
263
264         } else if (message instanceof FindLeader) {
265             getSender().tell(
266                 new FindLeaderReply(getLeaderAddress()),
267                 getSelf()
268             );
269         } else if (message instanceof GetOnDemandRaftState) {
270             onGetOnDemandRaftStats();
271         } else if (message instanceof InitiateCaptureSnapshot) {
272             captureSnapshot();
273         } else if (message instanceof SwitchBehavior) {
274             switchBehavior((SwitchBehavior) message);
275         } else if (message instanceof LeaderTransitioning) {
276             onLeaderTransitioning();
277         } else if (message instanceof Shutdown) {
278             onShutDown();
279         } else if (message instanceof Runnable) {
280             ((Runnable)message).run();
281         } else if (message instanceof NoopPayload) {
282             persistData(null, null, (NoopPayload)message);
283         } else if (!possiblyHandleBehaviorMessage(message)) {
284             handleNonRaftCommand(message);
285         }
286     }
287
288     private boolean possiblyHandleBehaviorMessage(final Object message) {
289         final RaftActorBehavior currentBehavior = getCurrentBehavior();
290         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
291
292         // A behavior indicates that it processed the change by returning a reference to the next behavior
293         // to be used. A null return indicates it has not processed the message and we should be passing it to
294         // the subclass for handling.
295         final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
296         if (nextBehavior != null) {
297             switchBehavior(state, nextBehavior);
298             return true;
299         }
300
301         return false;
302     }
303
304     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
305         LOG.debug("{}: Initiating leader transfer", persistenceId());
306
307         if (leadershipTransferInProgress == null) {
308             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
309             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
310                 @Override
311                 public void onSuccess(ActorRef raftActorRef) {
312                     leadershipTransferInProgress = null;
313                 }
314
315                 @Override
316                 public void onFailure(ActorRef raftActorRef) {
317                     leadershipTransferInProgress = null;
318                 }
319             });
320
321             leadershipTransferInProgress.addOnComplete(onComplete);
322             leadershipTransferInProgress.init();
323         } else {
324             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
325             leadershipTransferInProgress.addOnComplete(onComplete);
326         }
327     }
328
329     private void onShutDown() {
330         LOG.debug("{}: onShutDown", persistenceId());
331
332         if (shuttingDown) {
333             return;
334         }
335
336         shuttingDown = true;
337
338         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
339         if (currentBehavior.state() != RaftState.Leader) {
340             // For non-leaders shutdown is a no-op
341             self().tell(PoisonPill.getInstance(), self());
342             return;
343         }
344
345         if (context.hasFollowers()) {
346             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
347                 @Override
348                 public void onSuccess(ActorRef raftActorRef) {
349                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
350                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
351                 }
352
353                 @Override
354                 public void onFailure(ActorRef raftActorRef) {
355                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
356                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
357                 }
358             });
359         } else {
360             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
361                 @Override
362                 protected void doRun() {
363                     self().tell(PoisonPill.getInstance(), self());
364                 }
365
366                 @Override
367                 protected void doCancel() {
368                     self().tell(PoisonPill.getInstance(), self());
369                 }
370             });
371         }
372     }
373
374     private void onLeaderTransitioning() {
375         LOG.debug("{}: onLeaderTransitioning", persistenceId());
376         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
377         if (getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
378             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
379                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
380         }
381     }
382
383     private void switchBehavior(SwitchBehavior message) {
384         if (!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
385             RaftState newState = message.getNewState();
386             if ( newState == RaftState.Leader || newState == RaftState.Follower) {
387                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
388                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
389                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
390             } else {
391                 LOG.warn("Switching to behavior : {} - not supported", newState);
392             }
393         }
394     }
395
396     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
397         setCurrentBehavior(nextBehavior);
398         handleBehaviorChange(oldBehaviorState, nextBehavior);
399     }
400
401     @VisibleForTesting
402     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
403         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
404     }
405
406     private void onGetOnDemandRaftStats() {
407         // Debugging message to retrieve raft stats.
408
409         Map<String, String> peerAddresses = new HashMap<>();
410         Map<String, Boolean> peerVotingStates = new HashMap<>();
411         for (PeerInfo info: context.getPeers()) {
412             peerVotingStates.put(info.getId(), info.isVoting());
413             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
414         }
415
416         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
417         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
418                 .commitIndex(context.getCommitIndex())
419                 .currentTerm(context.getTermInformation().getCurrentTerm())
420                 .inMemoryJournalDataSize(replicatedLog().dataSize())
421                 .inMemoryJournalLogSize(replicatedLog().size())
422                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
423                 .lastApplied(context.getLastApplied())
424                 .lastIndex(replicatedLog().lastIndex())
425                 .lastTerm(replicatedLog().lastTerm())
426                 .leader(getLeaderId())
427                 .raftState(currentBehavior.state().toString())
428                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
429                 .snapshotIndex(replicatedLog().getSnapshotIndex())
430                 .snapshotTerm(replicatedLog().getSnapshotTerm())
431                 .votedFor(context.getTermInformation().getVotedFor())
432                 .isVoting(context.isVotingMember())
433                 .peerAddresses(peerAddresses)
434                 .peerVotingStates(peerVotingStates)
435                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
436
437         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
438         if (lastLogEntry != null) {
439             builder.lastLogIndex(lastLogEntry.getIndex());
440             builder.lastLogTerm(lastLogEntry.getTerm());
441         }
442
443         if (getCurrentBehavior() instanceof AbstractLeader) {
444             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
445             Collection<String> followerIds = leader.getFollowerIds();
446             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
447             for (String id: followerIds) {
448                 final FollowerLogInformation info = leader.getFollower(id);
449                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
450                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
451                         context.getPeerInfo(info.getId()).isVoting()));
452             }
453
454             builder.followerInfoList(followerInfoList);
455         }
456
457         sender().tell(builder.build(), self());
458
459     }
460
461     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
462         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
463
464         if (oldBehavior != currentBehavior){
465             onStateChanged();
466         }
467
468         String lastLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastLeaderId();
469         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
470         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
471
472         // it can happen that the state has not changed but the leader has changed.
473         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
474         if (!Objects.equals(lastLeaderId, currentBehavior.getLeaderId())
475                 || oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
476             if (roleChangeNotifier.isPresent()) {
477                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
478                         currentBehavior.getLeaderPayloadVersion()), getSelf());
479             }
480
481             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
482
483             if (leadershipTransferInProgress != null) {
484                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
485             }
486
487             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
488         }
489
490         if (roleChangeNotifier.isPresent()
491                 && (oldBehavior == null || oldBehavior.state() != currentBehavior.state())) {
492             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
493                     currentBehavior.state().name()), getSelf());
494         }
495     }
496
497     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
498         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
499     }
500
501     @Override
502     public long snapshotSequenceNr() {
503         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
504         // so that we can delete the persistent journal based on the saved sequence-number
505         // However , when akka replays the journal during recovery, it replays it from the sequence number when the
506         // snapshot was saved and not the number we saved. We would want to override it , by asking akka to use the
507         // last-sequence number known to us.
508         return context.getSnapshotManager().getLastSequenceNumber();
509     }
510
511     /**
512      * When a derived RaftActor needs to persist something it must call
513      * persistData.
514      */
515     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
516
517         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
518             context.getReplicatedLog().lastIndex() + 1,
519             context.getTermInformation().getCurrentTerm(), data);
520
521         LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
522
523         final RaftActorContext raftContext = getRaftActorContext();
524
525         replicatedLog().appendAndPersist(replicatedLogEntry, replicatedLogEntry1 -> {
526             if (!hasFollowers()) {
527                 // Increment the Commit Index and the Last Applied values
528                 raftContext.setCommitIndex(replicatedLogEntry1.getIndex());
529                 raftContext.setLastApplied(replicatedLogEntry1.getIndex());
530
531                 // Apply the state immediately.
532                 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry1), self());
533
534                 // Send a ApplyJournalEntries message so that we write the fact that we applied
535                 // the state to durable storage
536                 self().tell(new ApplyJournalEntries(replicatedLogEntry1.getIndex()), self());
537
538             } else {
539                 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry1);
540
541                 // Send message for replication
542                 getCurrentBehavior().handleMessage(getSelf(),
543                         new Replicate(clientActor, identifier, replicatedLogEntry1));
544             }
545         });
546     }
547
548     private ReplicatedLog replicatedLog() {
549         return context.getReplicatedLog();
550     }
551
552     protected String getId() {
553         return context.getId();
554     }
555
556     @VisibleForTesting
557     void setCurrentBehavior(RaftActorBehavior behavior) {
558         context.setCurrentBehavior(behavior);
559     }
560
561     protected RaftActorBehavior getCurrentBehavior() {
562         return context.getCurrentBehavior();
563     }
564
565     /**
566      * Derived actors can call the isLeader method to check if the current
567      * RaftActor is the Leader or not.
568      *
569      * @return true it this RaftActor is a Leader false otherwise
570      */
571     protected boolean isLeader() {
572         return context.getId().equals(getCurrentBehavior().getLeaderId());
573     }
574
575     protected final boolean isLeaderActive() {
576         return getRaftState() != RaftState.IsolatedLeader && getRaftState() != RaftState.PreLeader
577                 && !shuttingDown && !isLeadershipTransferInProgress();
578     }
579
580     private boolean isLeadershipTransferInProgress() {
581         return leadershipTransferInProgress != null && leadershipTransferInProgress.isTransferring();
582     }
583
584     /**
585      * Derived actor can call getLeader if they need a reference to the Leader.
586      * This would be useful for example in forwarding a request to an actor
587      * which is the leader
588      *
589      * @return A reference to the leader if known, null otherwise
590      */
591     public ActorSelection getLeader() {
592         String leaderAddress = getLeaderAddress();
593
594         if (leaderAddress == null) {
595             return null;
596         }
597
598         return context.actorSelection(leaderAddress);
599     }
600
601     /**
602      * Returns the id of the current leader.
603      *
604      * @return the current leader's id
605      */
606     protected final String getLeaderId() {
607         return getCurrentBehavior().getLeaderId();
608     }
609
610     @VisibleForTesting
611     protected final RaftState getRaftState() {
612         return getCurrentBehavior().state();
613     }
614
615     protected Long getCurrentTerm() {
616         return context.getTermInformation().getCurrentTerm();
617     }
618
619     protected RaftActorContext getRaftActorContext() {
620         return context;
621     }
622
623     protected void updateConfigParams(ConfigParams configParams) {
624
625         // obtain the RaftPolicy for oldConfigParams and the updated one.
626         String oldRaftPolicy = context.getConfigParams().getCustomRaftPolicyImplementationClass();
627         String newRaftPolicy = configParams.getCustomRaftPolicyImplementationClass();
628
629         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
630             oldRaftPolicy, newRaftPolicy);
631         context.setConfigParams(configParams);
632         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
633             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
634             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
635             // avoids potential disruption. Otherwise, switch to Follower normally.
636             RaftActorBehavior behavior = getCurrentBehavior();
637             if (behavior != null && behavior.state() == RaftState.Follower) {
638                 String previousLeaderId = behavior.getLeaderId();
639                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
640
641                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(),
642                         previousLeaderId);
643
644                 changeCurrentBehavior(new Follower(context, previousLeaderId, previousLeaderPayloadVersion));
645             } else {
646                 initializeBehavior();
647             }
648         }
649     }
650
651     public final DataPersistenceProvider persistence() {
652         return delegatingPersistenceProvider.getDelegate();
653     }
654
655     public void setPersistence(DataPersistenceProvider provider) {
656         delegatingPersistenceProvider.setDelegate(provider);
657     }
658
659     protected void setPersistence(boolean persistent) {
660         DataPersistenceProvider currentPersistence = persistence();
661         if (persistent && (currentPersistence == null || !currentPersistence.isRecoveryApplicable())) {
662             setPersistence(new PersistentDataProvider(this));
663
664             if (getCurrentBehavior() != null) {
665                 LOG.info("{}: Persistence has been enabled - capturing snapshot", persistenceId());
666                 captureSnapshot();
667             }
668         } else if (!persistent && (currentPersistence == null || currentPersistence.isRecoveryApplicable())) {
669             setPersistence(new NonPersistentDataProvider() {
670                 /**
671                  * The way snapshotting works is,
672                  * <ol>
673                  * <li> RaftActor calls createSnapshot on the Shard
674                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
675                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
676                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
677                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
678                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
679                  * </ol>
680                  */
681                 @Override
682                 public void saveSnapshot(Object object) {
683                     // Make saving Snapshot successful
684                     // Committing the snapshot here would end up calling commit in the creating state which would
685                     // be a state violation. That's why now we send a message to commit the snapshot.
686                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
687                 }
688             });
689         }
690     }
691
692     /**
693      * setPeerAddress sets the address of a known peer at a later time.
694      *
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      *
701      * <p>
702      * Note that if the peerId does not match the list of peers passed to
703      * this actor during construction an IllegalStateException will be thrown.
704      */
705     protected void setPeerAddress(String peerId, String peerAddress) {
706         context.setPeerAddress(peerId, peerAddress);
707     }
708
709     /**
710      * The applyState method will be called by the RaftActor when some data
711      * needs to be applied to the actor's state.
712      *
713      * @param clientActor A reference to the client who sent this message. This
714      *                    is the same reference that was passed to persistData
715      *                    by the derived actor. clientActor may be null when
716      *                    the RaftActor is behaving as a follower or during
717      *                    recovery.
718      * @param identifier  The identifier of the persisted data. This is also
719      *                    the same identifier that was passed to persistData by
720      *                    the derived actor. identifier may be null when
721      *                    the RaftActor is behaving as a follower or during
722      *                    recovery
723      * @param data        A piece of data that was persisted by the persistData call.
724      *                    This should NEVER be null.
725      */
726     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
727
728     /**
729      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
730      */
731     @Nonnull
732     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
733
734     /**
735      * This method is called when recovery is complete.
736      */
737     protected abstract void onRecoveryComplete();
738
739     /**
740      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
741      */
742     @Nonnull
743     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
744
745     /**
746      * This method will be called by the RaftActor when the state of the
747      * RaftActor changes. The derived actor can then use methods like
748      * isLeader or getLeader to do something useful
749      */
750     protected abstract void onStateChanged();
751
752     /**
753      * Notifier Actor for this RaftActor to notify when a role change happens.
754      *
755      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
756      */
757     protected abstract Optional<ActorRef> getRoleChangeNotifier();
758
759     /**
760      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
761      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
762      * work prior to performing the operation. On completion of any work, the run method must be called on the
763      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
764      * this actor's thread dispatcher as as it modifies internal state.
765      *
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     private String getLeaderAddress() {
779         if (isLeader()) {
780             return getSelf().path().toString();
781         }
782         String leaderId = getLeaderId();
783         if (leaderId == null) {
784             return null;
785         }
786         String peerAddress = context.getPeerAddress(leaderId);
787         LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}", persistenceId(), leaderId, peerAddress);
788
789         return peerAddress;
790     }
791
792     protected boolean hasFollowers() {
793         return getRaftActorContext().hasFollowers();
794     }
795
796     private void captureSnapshot() {
797         SnapshotManager snapshotManager = context.getSnapshotManager();
798
799         if (!snapshotManager.isCapturing()) {
800             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
801             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
802                 replicatedLog().last(), idx);
803
804             snapshotManager.capture(replicatedLog().last(), idx);
805         }
806     }
807
808     /**
809      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
810      * in which case we need to step down.
811      */
812     void becomeNonVoting() {
813         if (isLeader()) {
814             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
815                 @Override
816                 public void onSuccess(ActorRef raftActorRef) {
817                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
818                     ensureFollowerState();
819                 }
820
821                 @Override
822                 public void onFailure(ActorRef raftActorRef) {
823                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
824                     ensureFollowerState();
825                 }
826
827                 private void ensureFollowerState() {
828                     // Whether or not leadership transfer succeeded, we have to step down as leader and
829                     // switch to Follower so ensure that.
830                     if (getRaftState() != RaftState.Follower) {
831                         initializeBehavior();
832                     }
833                 }
834             });
835         }
836     }
837
838     /**
839      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
840      */
841     private abstract static class BehaviorState implements Immutable {
842         @Nullable abstract RaftActorBehavior getBehavior();
843
844         @Nullable abstract String getLastValidLeaderId();
845
846         @Nullable abstract String getLastLeaderId();
847
848         @Nullable abstract short getLeaderPayloadVersion();
849     }
850
851     /**
852      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
853      */
854     private static final class SimpleBehaviorState extends BehaviorState {
855         private final RaftActorBehavior behavior;
856         private final String lastValidLeaderId;
857         private final String lastLeaderId;
858         private final short leaderPayloadVersion;
859
860         SimpleBehaviorState(final String lastValidLeaderId, final String lastLeaderId,
861                 final RaftActorBehavior behavior) {
862             this.lastValidLeaderId = lastValidLeaderId;
863             this.lastLeaderId = lastLeaderId;
864             this.behavior = Preconditions.checkNotNull(behavior);
865             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
866         }
867
868         @Override
869         RaftActorBehavior getBehavior() {
870             return behavior;
871         }
872
873         @Override
874         String getLastValidLeaderId() {
875             return lastValidLeaderId;
876         }
877
878         @Override
879         short getLeaderPayloadVersion() {
880             return leaderPayloadVersion;
881         }
882
883         @Override
884         String getLastLeaderId() {
885             return lastLeaderId;
886         }
887     }
888
889     /**
890      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
891      * An instance is created for each RaftActor. It has two functions:
892      * - it keeps track of the last leader ID we have encountered since we have been created
893      * - it creates state capture needed to transition from one behavior to the next
894      */
895     private static final class BehaviorStateTracker {
896         /**
897          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
898          * allowed before we receive the first message, we know the leader ID to be null.
899          */
900         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
901             @Override
902             RaftActorBehavior getBehavior() {
903                 return null;
904             }
905
906             @Override
907             String getLastValidLeaderId() {
908                 return null;
909             }
910
911             @Override
912             short getLeaderPayloadVersion() {
913                 return -1;
914             }
915
916             @Override
917             String getLastLeaderId() {
918                 return null;
919             }
920         };
921
922         private String lastValidLeaderId;
923         private String lastLeaderId;
924
925         BehaviorState capture(final RaftActorBehavior behavior) {
926             if (behavior == null) {
927                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
928                 return NULL_BEHAVIOR_STATE;
929             }
930
931             lastLeaderId = behavior.getLeaderId();
932             if (lastLeaderId != null) {
933                 lastValidLeaderId = lastLeaderId;
934             }
935
936             return new SimpleBehaviorState(lastValidLeaderId, lastLeaderId, behavior);
937         }
938     }
939
940 }