Remove snapshot after startup and fix related bug
[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     }
175
176     protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
177         return new RaftActorRecoverySupport(context, getRaftActorRecoveryCohort());
178     }
179
180     @VisibleForTesting
181     void initializeBehavior(){
182         changeCurrentBehavior(new Follower(context));
183     }
184
185     @VisibleForTesting
186     protected void changeCurrentBehavior(RaftActorBehavior newBehavior) {
187         final RaftActorBehavior currentBehavior = getCurrentBehavior();
188         if (currentBehavior != null) {
189             try {
190                 currentBehavior.close();
191             } catch (Exception e) {
192                 LOG.warn("{}: Error closing behavior {}", persistence(), currentBehavior, e);
193             }
194         }
195
196         final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
197         setCurrentBehavior(newBehavior);
198         handleBehaviorChange(state, newBehavior);
199     }
200
201     /**
202      * Method exposed for subclasses to plug-in their logic. This method is invoked by {@link #handleCommand(Object)}
203      * for messages which are not handled by this class. Subclasses overriding this class should fall back to this
204      * implementation for messages which they do not handle
205      *
206      * @param message Incoming command message
207      */
208     protected void handleNonRaftCommand(final Object message) {
209         unhandled(message);
210     }
211
212     /**
213      * @deprecated This method is not final for testing purposes. DO NOT OVERRIDE IT, override
214      * {@link #handleNonRaftCommand(Object)} instead.
215      */
216     @Deprecated
217     @Override
218     // FIXME: make this method final once our unit tests do not need to override it
219     protected void handleCommand(final Object message) {
220         if (serverConfigurationSupport.handleMessage(message, getSender())) {
221             return;
222         }
223         if (snapshotSupport.handleSnapshotMessage(message, getSender())) {
224             return;
225         }
226
227         if (message instanceof ApplyState) {
228             ApplyState applyState = (ApplyState) message;
229
230             long startTime = System.nanoTime();
231
232             if(LOG.isDebugEnabled()) {
233                 LOG.debug("{}: Applying state for log index {} data {}",
234                     persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
235                     applyState.getReplicatedLogEntry().getData());
236             }
237
238             applyState(applyState.getClientActor(), applyState.getIdentifier(),
239                 applyState.getReplicatedLogEntry().getData());
240
241             long elapsedTime = System.nanoTime() - startTime;
242             if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
243                 LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
244                         TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
245             }
246
247             if (!hasFollowers()) {
248                 // for single node, the capture should happen after the apply state
249                 // as we delete messages from the persistent journal which have made it to the snapshot
250                 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
251                 // and recovery shows data missing
252                 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
253
254                 context.getSnapshotManager().trimLog(context.getLastApplied());
255             }
256
257         } else if (message instanceof ApplyJournalEntries) {
258             ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
259             if(LOG.isDebugEnabled()) {
260                 LOG.debug("{}: Persisting ApplyJournalEntries with index={}", persistenceId(), applyEntries.getToIndex());
261             }
262
263             persistence().persist(applyEntries, NoopProcedure.instance());
264
265         } else if (message instanceof FindLeader) {
266             getSender().tell(
267                 new FindLeaderReply(getLeaderAddress()),
268                 getSelf()
269             );
270         } else if(message instanceof GetOnDemandRaftState) {
271             onGetOnDemandRaftStats();
272         } else if(message instanceof InitiateCaptureSnapshot) {
273             captureSnapshot();
274         } else if(message instanceof SwitchBehavior) {
275             switchBehavior(((SwitchBehavior) message));
276         } else if(message instanceof LeaderTransitioning) {
277             onLeaderTransitioning();
278         } else if(message instanceof Shutdown) {
279             onShutDown();
280         } else if(message instanceof Runnable) {
281             ((Runnable)message).run();
282         } else {
283             // Processing the message may affect the state, hence we need to capture it
284             final RaftActorBehavior currentBehavior = getCurrentBehavior();
285             final BehaviorState state = behaviorStateTracker.capture(currentBehavior);
286
287             // A behavior indicates that it processed the change by returning a reference to the next behavior
288             // to be used. A null return indicates it has not processed the message and we should be passing it to
289             // the subclass for handling.
290             final RaftActorBehavior nextBehavior = currentBehavior.handleMessage(getSender(), message);
291             if (nextBehavior != null) {
292                 switchBehavior(state, nextBehavior);
293             } else {
294                 handleNonRaftCommand(message);
295             }
296         }
297     }
298
299     private void initiateLeadershipTransfer(final RaftActorLeadershipTransferCohort.OnComplete onComplete) {
300         LOG.debug("{}: Initiating leader transfer", persistenceId());
301
302         if(leadershipTransferInProgress == null) {
303             leadershipTransferInProgress = new RaftActorLeadershipTransferCohort(this);
304             leadershipTransferInProgress.addOnComplete(new RaftActorLeadershipTransferCohort.OnComplete() {
305                 @Override
306                 public void onSuccess(ActorRef raftActorRef) {
307                     leadershipTransferInProgress = null;
308                 }
309
310                 @Override
311                 public void onFailure(ActorRef raftActorRef) {
312                     leadershipTransferInProgress = null;
313                 }
314             });
315
316             leadershipTransferInProgress.addOnComplete(onComplete);
317             leadershipTransferInProgress.init();
318         } else {
319             LOG.debug("{}: prior leader transfer in progress - adding callback", persistenceId());
320             leadershipTransferInProgress.addOnComplete(onComplete);
321         }
322     }
323
324     private void onShutDown() {
325         LOG.debug("{}: onShutDown", persistenceId());
326
327         if(shuttingDown) {
328             return;
329         }
330
331         shuttingDown = true;
332
333         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
334         if (currentBehavior.state() != RaftState.Leader) {
335             // For non-leaders shutdown is a no-op
336             self().tell(PoisonPill.getInstance(), self());
337             return;
338         }
339
340         if (context.hasFollowers()) {
341             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
342                 @Override
343                 public void onSuccess(ActorRef raftActorRef) {
344                     LOG.debug("{}: leader transfer succeeded - sending PoisonPill", persistenceId());
345                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
346                 }
347
348                 @Override
349                 public void onFailure(ActorRef raftActorRef) {
350                     LOG.debug("{}: leader transfer failed - sending PoisonPill", persistenceId());
351                     raftActorRef.tell(PoisonPill.getInstance(), raftActorRef);
352                 }
353             });
354         } else {
355             pauseLeader(new TimedRunnable(context.getConfigParams().getElectionTimeOutInterval(), this) {
356                 @Override
357                 protected void doRun() {
358                     self().tell(PoisonPill.getInstance(), self());
359                 }
360
361                 @Override
362                 protected void doCancel() {
363                     self().tell(PoisonPill.getInstance(), self());
364                 }
365             });
366         }
367     }
368
369     private void onLeaderTransitioning() {
370         LOG.debug("{}: onLeaderTransitioning", persistenceId());
371         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
372         if(getRaftState() == RaftState.Follower && roleChangeNotifier.isPresent()) {
373             roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), null,
374                 getCurrentBehavior().getLeaderPayloadVersion()), getSelf());
375         }
376     }
377
378     private void switchBehavior(SwitchBehavior message) {
379         if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) {
380             RaftState newState = message.getNewState();
381             if( newState == RaftState.Leader || newState == RaftState.Follower) {
382                 switchBehavior(behaviorStateTracker.capture(getCurrentBehavior()),
383                     AbstractRaftActorBehavior.createBehavior(context, message.getNewState()));
384                 getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), "");
385             } else {
386                 LOG.warn("Switching to behavior : {} - not supported", newState);
387             }
388         }
389     }
390
391     private void switchBehavior(final BehaviorState oldBehaviorState, final RaftActorBehavior nextBehavior) {
392         setCurrentBehavior(nextBehavior);
393         handleBehaviorChange(oldBehaviorState, nextBehavior);
394     }
395
396     @VisibleForTesting
397     RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
398         return new RaftActorSnapshotMessageSupport(context, getRaftActorSnapshotCohort());
399     }
400
401     private void onGetOnDemandRaftStats() {
402         // Debugging message to retrieve raft stats.
403
404         Map<String, String> peerAddresses = new HashMap<>();
405         Map<String, Boolean> peerVotingStates = new HashMap<>();
406         for(PeerInfo info: context.getPeers()) {
407             peerVotingStates.put(info.getId(), info.isVoting());
408             peerAddresses.put(info.getId(), info.getAddress() != null ? info.getAddress() : "");
409         }
410
411         final RaftActorBehavior currentBehavior = context.getCurrentBehavior();
412         OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
413                 .commitIndex(context.getCommitIndex())
414                 .currentTerm(context.getTermInformation().getCurrentTerm())
415                 .inMemoryJournalDataSize(replicatedLog().dataSize())
416                 .inMemoryJournalLogSize(replicatedLog().size())
417                 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
418                 .lastApplied(context.getLastApplied())
419                 .lastIndex(replicatedLog().lastIndex())
420                 .lastTerm(replicatedLog().lastTerm())
421                 .leader(getLeaderId())
422                 .raftState(currentBehavior.state().toString())
423                 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
424                 .snapshotIndex(replicatedLog().getSnapshotIndex())
425                 .snapshotTerm(replicatedLog().getSnapshotTerm())
426                 .votedFor(context.getTermInformation().getVotedFor())
427                 .isVoting(context.isVotingMember())
428                 .peerAddresses(peerAddresses)
429                 .peerVotingStates(peerVotingStates)
430                 .customRaftPolicyClassName(context.getConfigParams().getCustomRaftPolicyImplementationClass());
431
432         ReplicatedLogEntry lastLogEntry = replicatedLog().last();
433         if (lastLogEntry != null) {
434             builder.lastLogIndex(lastLogEntry.getIndex());
435             builder.lastLogTerm(lastLogEntry.getTerm());
436         }
437
438         if(getCurrentBehavior() instanceof AbstractLeader) {
439             AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
440             Collection<String> followerIds = leader.getFollowerIds();
441             List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
442             for(String id: followerIds) {
443                 final FollowerLogInformation info = leader.getFollower(id);
444                 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
445                         info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()),
446                         context.getPeerInfo(info.getId()).isVoting()));
447             }
448
449             builder.followerInfoList(followerInfoList);
450         }
451
452         sender().tell(builder.build(), self());
453
454     }
455
456     private void handleBehaviorChange(BehaviorState oldBehaviorState, RaftActorBehavior currentBehavior) {
457         RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
458
459         if (oldBehavior != currentBehavior){
460             onStateChanged();
461         }
462
463         String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId();
464         String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
465
466         // it can happen that the state has not changed but the leader has changed.
467         Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
468         if(!Objects.equals(lastValidLeaderId, currentBehavior.getLeaderId()) ||
469            oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) {
470             if(roleChangeNotifier.isPresent()) {
471                 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(),
472                         currentBehavior.getLeaderPayloadVersion()), getSelf());
473             }
474
475             onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId());
476
477             if(leadershipTransferInProgress != null) {
478                 leadershipTransferInProgress.onNewLeader(currentBehavior.getLeaderId());
479             }
480
481             serverConfigurationSupport.onNewLeader(currentBehavior.getLeaderId());
482         }
483
484         if (roleChangeNotifier.isPresent() &&
485                 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
486             roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
487                     currentBehavior.state().name()), getSelf());
488         }
489     }
490
491     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
492         return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
493     }
494
495     @Override
496     public long snapshotSequenceNr() {
497         // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
498         // so that we can delete the persistent journal based on the saved sequence-number
499         // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
500         // was saved and not the number we saved.
501         // We would want to override it , by asking akka to use the last-sequence number known to us.
502         return context.getSnapshotManager().getLastSequenceNumber();
503     }
504
505     /**
506      * When a derived RaftActor needs to persist something it must call
507      * persistData.
508      *
509      * @param clientActor
510      * @param identifier
511      * @param data
512      */
513     protected final void persistData(final ActorRef clientActor, final Identifier identifier, final Payload data) {
514
515         ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
516             context.getReplicatedLog().lastIndex() + 1,
517             context.getTermInformation().getCurrentTerm(), data);
518
519         if(LOG.isDebugEnabled()) {
520             LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
521         }
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 if (clientActor != null) {
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 && !shuttingDown &&
577                 !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     protected 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      *
603      * @return the current leader's id
604      */
605     protected final String getLeaderId(){
606         return getCurrentBehavior().getLeaderId();
607     }
608
609     @VisibleForTesting
610     protected final RaftState getRaftState() {
611         return getCurrentBehavior().state();
612     }
613
614     protected Long getCurrentTerm(){
615         return context.getTermInformation().getCurrentTerm();
616     }
617
618     protected RaftActorContext getRaftActorContext() {
619         return context;
620     }
621
622     protected void updateConfigParams(ConfigParams configParams) {
623
624         // obtain the RaftPolicy for oldConfigParams and the updated one.
625         String oldRaftPolicy = context.getConfigParams().
626             getCustomRaftPolicyImplementationClass();
627         String newRaftPolicy = configParams.
628             getCustomRaftPolicyImplementationClass();
629
630         LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(),
631             oldRaftPolicy, newRaftPolicy);
632         context.setConfigParams(configParams);
633         if (!Objects.equals(oldRaftPolicy, newRaftPolicy)) {
634             // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower
635             // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This
636             // avoids potential disruption. Otherwise, switch to Follower normally.
637             RaftActorBehavior behavior = getCurrentBehavior();
638             if (behavior != null && behavior.state() == RaftState.Follower) {
639                 String previousLeaderId = behavior.getLeaderId();
640                 short previousLeaderPayloadVersion = behavior.getLeaderPayloadVersion();
641
642                 LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), 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         if(persistent) {
661             setPersistence(new PersistentDataProvider(this));
662         } else {
663             setPersistence(new NonPersistentDataProvider() {
664                 /**
665                  * The way snapshotting works is,
666                  * <ol>
667                  * <li> RaftActor calls createSnapshot on the Shard
668                  * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
669                  * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
670                  * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
671                  * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
672                  * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
673                  * </ol>
674                  */
675                 @Override
676                 public void saveSnapshot(Object o) {
677                     // Make saving Snapshot successful
678                     // Committing the snapshot here would end up calling commit in the creating state which would
679                     // be a state violation. That's why now we send a message to commit the snapshot.
680                     self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
681                 }
682             });
683         }
684     }
685
686     /**
687      * setPeerAddress sets the address of a known peer at a later time.
688      * <p>
689      * This is to account for situations where a we know that a peer
690      * exists but we do not know an address up-front. This may also be used in
691      * situations where a known peer starts off in a different location and we
692      * need to change it's address
693      * <p>
694      * Note that if the peerId does not match the list of peers passed to
695      * this actor during construction an IllegalStateException will be thrown.
696      *
697      * @param peerId
698      * @param peerAddress
699      */
700     protected void setPeerAddress(String peerId, String peerAddress){
701         context.setPeerAddress(peerId, peerAddress);
702     }
703
704     /**
705      * The applyState method will be called by the RaftActor when some data
706      * needs to be applied to the actor's state
707      *
708      * @param clientActor A reference to the client who sent this message. This
709      *                    is the same reference that was passed to persistData
710      *                    by the derived actor. clientActor may be null when
711      *                    the RaftActor is behaving as a follower or during
712      *                    recovery.
713      * @param identifier  The identifier of the persisted data. This is also
714      *                    the same identifier that was passed to persistData by
715      *                    the derived actor. identifier may be null when
716      *                    the RaftActor is behaving as a follower or during
717      *                    recovery
718      * @param data        A piece of data that was persisted by the persistData call.
719      *                    This should NEVER be null.
720      */
721     protected abstract void applyState(ActorRef clientActor, Identifier identifier, Object data);
722
723     /**
724      * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
725      */
726     @Nonnull
727     protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
728
729     /**
730      * This method is called when recovery is complete.
731      */
732     protected abstract void onRecoveryComplete();
733
734     /**
735      * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
736      */
737     @Nonnull
738     protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
739
740     /**
741      * This method will be called by the RaftActor when the state of the
742      * RaftActor changes. The derived actor can then use methods like
743      * isLeader or getLeader to do something useful
744      */
745     protected abstract void onStateChanged();
746
747     /**
748      * Notifier Actor for this RaftActor to notify when a role change happens
749      * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
750      */
751     protected abstract Optional<ActorRef> getRoleChangeNotifier();
752
753     /**
754      * This method is called prior to operations such as leadership transfer and actor shutdown when the leader
755      * must pause or stop its duties. This method allows derived classes to gracefully pause or finish current
756      * work prior to performing the operation. On completion of any work, the run method must be called on the
757      * given Runnable to proceed with the given operation. <b>Important:</b> the run method must be called on
758      * this actor's thread dispatcher as as it modifies internal state.
759      * <p>
760      * The default implementation immediately runs the operation.
761      *
762      * @param operation the operation to run
763      */
764     protected void pauseLeader(Runnable operation) {
765         operation.run();
766     }
767
768     protected void onLeaderChanged(String oldLeader, String newLeader) {
769
770     };
771
772     private String getLeaderAddress(){
773         if(isLeader()){
774             return getSelf().path().toString();
775         }
776         String leaderId = getLeaderId();
777         if (leaderId == null) {
778             return null;
779         }
780         String peerAddress = context.getPeerAddress(leaderId);
781         if(LOG.isDebugEnabled()) {
782             LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
783                     persistenceId(), leaderId, peerAddress);
784         }
785
786         return peerAddress;
787     }
788
789     protected boolean hasFollowers(){
790         return getRaftActorContext().hasFollowers();
791     }
792
793     private void captureSnapshot() {
794         SnapshotManager snapshotManager = context.getSnapshotManager();
795
796         if (!snapshotManager.isCapturing()) {
797             final long idx = getCurrentBehavior().getReplicatedToAllIndex();
798             LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}",
799                 replicatedLog().last(), idx);
800
801             snapshotManager.capture(replicatedLog().last(), idx);
802         }
803     }
804
805     /**
806      * Switch this member to non-voting status. This is a no-op for all behaviors except when we are the leader,
807      * in which case we need to step down.
808      */
809     void becomeNonVoting() {
810         if (isLeader()) {
811             initiateLeadershipTransfer(new RaftActorLeadershipTransferCohort.OnComplete() {
812                 @Override
813                 public void onSuccess(ActorRef raftActorRef) {
814                     LOG.debug("{}: leader transfer succeeded after change to non-voting", persistenceId());
815                     ensureFollowerState();
816                 }
817
818                 @Override
819                 public void onFailure(ActorRef raftActorRef) {
820                     LOG.debug("{}: leader transfer failed after change to non-voting", persistenceId());
821                     ensureFollowerState();
822                 }
823
824                 private void ensureFollowerState() {
825                     // Whether or not leadership transfer succeeded, we have to step down as leader and
826                     // switch to Follower so ensure that.
827                     if (getRaftState() != RaftState.Follower) {
828                         initializeBehavior();
829                     }
830                 }
831             });
832         }
833     }
834
835     /**
836      * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
837      *             whose type for fromIndex is long instead of int. This class was kept for backwards
838      *             compatibility with Helium.
839      */
840     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
841     @SuppressWarnings("serial")
842     @Deprecated
843     static class DeleteEntries implements Serializable {
844         private final int fromIndex;
845
846         public DeleteEntries(int fromIndex) {
847             this.fromIndex = fromIndex;
848         }
849
850         public int getFromIndex() {
851             return fromIndex;
852         }
853     }
854
855     /**
856      * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
857      *             which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
858      */
859     // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
860     @SuppressWarnings("serial")
861     @Deprecated
862     static class UpdateElectionTerm implements Serializable {
863         private final long currentTerm;
864         private final String votedFor;
865
866         public UpdateElectionTerm(long currentTerm, String votedFor) {
867             this.currentTerm = currentTerm;
868             this.votedFor = votedFor;
869         }
870
871         public long getCurrentTerm() {
872             return currentTerm;
873         }
874
875         public String getVotedFor() {
876             return votedFor;
877         }
878     }
879
880     /**
881      * A point-in-time capture of {@link RaftActorBehavior} state critical for transitioning between behaviors.
882      */
883     private static abstract class BehaviorState implements Immutable {
884         @Nullable abstract RaftActorBehavior getBehavior();
885         @Nullable abstract String getLastValidLeaderId();
886         @Nullable abstract short getLeaderPayloadVersion();
887     }
888
889     /**
890      * A {@link BehaviorState} corresponding to non-null {@link RaftActorBehavior} state.
891      */
892     private static final class SimpleBehaviorState extends BehaviorState {
893         private final RaftActorBehavior behavior;
894         private final String lastValidLeaderId;
895         private final short leaderPayloadVersion;
896
897         SimpleBehaviorState(final String lastValidLeaderId, final RaftActorBehavior behavior) {
898             this.lastValidLeaderId = lastValidLeaderId;
899             this.behavior = Preconditions.checkNotNull(behavior);
900             this.leaderPayloadVersion = behavior.getLeaderPayloadVersion();
901         }
902
903         @Override
904         RaftActorBehavior getBehavior() {
905             return behavior;
906         }
907
908         @Override
909         String getLastValidLeaderId() {
910             return lastValidLeaderId;
911         }
912
913         @Override
914         short getLeaderPayloadVersion() {
915             return leaderPayloadVersion;
916         }
917     }
918
919     /**
920      * Class tracking behavior-related information, which we need to keep around and pass across behavior switches.
921      * An instance is created for each RaftActor. It has two functions:
922      * - it keeps track of the last leader ID we have encountered since we have been created
923      * - it creates state capture needed to transition from one behavior to the next
924      */
925     private static final class BehaviorStateTracker {
926         /**
927          * A {@link BehaviorState} corresponding to null {@link RaftActorBehavior} state. Since null behavior is only
928          * allowed before we receive the first message, we know the leader ID to be null.
929          */
930         private static final BehaviorState NULL_BEHAVIOR_STATE = new BehaviorState() {
931             @Override
932             RaftActorBehavior getBehavior() {
933                 return null;
934             }
935
936             @Override
937             String getLastValidLeaderId() {
938                 return null;
939             }
940
941             @Override
942             short getLeaderPayloadVersion() {
943                 return -1;
944             }
945         };
946
947         private String lastValidLeaderId;
948
949         BehaviorState capture(final RaftActorBehavior behavior) {
950             if (behavior == null) {
951                 Verify.verify(lastValidLeaderId == null, "Null behavior with non-null last leader");
952                 return NULL_BEHAVIOR_STATE;
953             }
954
955             final String leaderId = behavior.getLeaderId();
956             if (leaderId != null) {
957                 lastValidLeaderId = leaderId;
958             }
959
960             return new SimpleBehaviorState(lastValidLeaderId, behavior);
961         }
962     }
963
964 }

©2013 OpenDaylight, A Linux Foundation Collaborative Project. All Rights Reserved.
OpenDaylight is a registered trademark of The OpenDaylight Project, Inc.
Linux Foundation and OpenDaylight are registered trademarks of the Linux Foundation.
Linux is a registered trademark of Linus Torvalds.