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

©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.