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.
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
10 package org.opendaylight.controller.cluster.raft;
12 import akka.actor.ActorRef;
13 import akka.actor.ActorSelection;
14 import akka.japi.Procedure;
15 import akka.persistence.SnapshotSelectionCriteria;
16 import com.google.common.annotations.VisibleForTesting;
17 import com.google.common.base.Objects;
18 import com.google.common.base.Optional;
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;
25 import java.util.concurrent.TimeUnit;
26 import javax.annotation.Nonnull;
27 import org.apache.commons.lang3.time.DurationFormatUtils;
28 import org.opendaylight.controller.cluster.DataPersistenceProvider;
29 import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider;
30 import org.opendaylight.controller.cluster.NonPersistentDataProvider;
31 import org.opendaylight.controller.cluster.PersistentDataProvider;
32 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
33 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
34 import org.opendaylight.controller.cluster.notifications.RoleChanged;
35 import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
36 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
37 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
38 import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
39 import org.opendaylight.controller.cluster.raft.behaviors.DelegatingRaftActorBehavior;
40 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
41 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
42 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
43 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
44 import org.opendaylight.controller.cluster.raft.client.messages.FollowerInfo;
45 import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState;
46 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
47 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
48 import org.slf4j.Logger;
49 import org.slf4j.LoggerFactory;
52 * RaftActor encapsulates a state machine that needs to be kept synchronized
53 * in a cluster. It implements the RAFT algorithm as described in the paper
54 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
55 * In Search of an Understandable Consensus Algorithm</a>
57 * RaftActor has 3 states and each state has a certain behavior associated
58 * with it. A Raft actor can behave as,
61 * <li> A Follower (or) </li>
62 * <li> A Candidate </li>
66 * A RaftActor MUST be a Leader in order to accept requests from clients to
67 * change the state of it's encapsulated state machine. Once a RaftActor becomes
68 * a Leader it is also responsible for ensuring that all followers ultimately
69 * have the same log and therefore the same state machine as itself.
72 * The current behavior of a RaftActor determines how election for leadership
73 * is initiated and how peer RaftActors react to request for votes.
76 * Each RaftActor also needs to know the current election term. It uses this
77 * information for a couple of things. One is to simply figure out who it
78 * voted for in the last election. Another is to figure out if the message
79 * it received to update it's state is stale.
82 * The RaftActor uses akka-persistence to store it's replicated log.
83 * Furthermore through it's behaviors a Raft Actor determines
86 * <li> when a log entry should be persisted </li>
87 * <li> when a log entry should be applied to the state machine (and) </li>
88 * <li> when a snapshot should be saved </li>
91 public abstract class RaftActor extends AbstractUntypedPersistentActor {
93 private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
95 protected final Logger LOG = LoggerFactory.getLogger(getClass());
98 * The current state determines the current behavior of a RaftActor
99 * A Raft Actor always starts off in the Follower State
101 private final DelegatingRaftActorBehavior currentBehavior = new DelegatingRaftActorBehavior();
104 * This context should NOT be passed directly to any other actor it is
105 * only to be consumed by the RaftActorBehaviors
107 private final RaftActorContextImpl context;
109 private final DelegatingPersistentDataProvider delegatingPersistenceProvider = new DelegatingPersistentDataProvider(null);
111 private RaftActorRecoverySupport raftRecovery;
113 private RaftActorSnapshotMessageSupport snapshotSupport;
115 private final BehaviorStateHolder reusableBehaviorStateHolder = new BehaviorStateHolder();
117 public RaftActor(String id, Map<String, String> peerAddresses,
118 Optional<ConfigParams> configParams, short payloadVersion) {
120 context = new RaftActorContextImpl(this.getSelf(),
121 this.getContext(), id, new ElectionTermImpl(delegatingPersistenceProvider, id, LOG),
122 -1, -1, peerAddresses,
123 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
124 delegatingPersistenceProvider, LOG);
126 context.setPayloadVersion(payloadVersion);
127 context.setReplicatedLog(ReplicatedLogImpl.newInstance(context, currentBehavior));
131 public void preStart() throws Exception {
132 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
133 context.getConfigParams().getJournalRecoveryLogBatchSize());
137 snapshotSupport = newRaftActorSnapshotMessageSupport();
141 public void postStop() {
142 if(currentBehavior.getDelegate() != null) {
144 currentBehavior.close();
145 } catch (Exception e) {
146 LOG.debug("{}: Error closing behavior {}", persistenceId(), currentBehavior.state());
154 public void handleRecover(Object message) {
155 if(raftRecovery == null) {
156 raftRecovery = newRaftActorRecoverySupport();
159 boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message);
160 if(recoveryComplete) {
161 if(!persistence().isRecoveryApplicable()) {
162 // Delete all the messages from the akka journal so that we do not end up with consistency issues
163 // Note I am not using the dataPersistenceProvider and directly using the akka api here
164 deleteMessages(lastSequenceNr());
166 // Delete all the akka snapshots as they will not be needed
167 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
170 onRecoveryComplete();
172 initializeBehavior();
178 protected RaftActorRecoverySupport newRaftActorRecoverySupport() {
179 return new RaftActorRecoverySupport(context, currentBehavior, getRaftActorRecoveryCohort());
182 protected void initializeBehavior(){
183 changeCurrentBehavior(new Follower(context));
186 protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
187 reusableBehaviorStateHolder.init(getCurrentBehavior());
188 setCurrentBehavior(newBehavior);
189 handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior());
193 public void handleCommand(Object message) {
194 if (message instanceof ApplyState){
195 ApplyState applyState = (ApplyState) message;
197 long elapsedTime = (System.nanoTime() - applyState.getStartTime());
198 if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
199 LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
200 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
203 if(LOG.isDebugEnabled()) {
204 LOG.debug("{}: Applying state for log index {} data {}",
205 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
206 applyState.getReplicatedLogEntry().getData());
209 applyState(applyState.getClientActor(), applyState.getIdentifier(),
210 applyState.getReplicatedLogEntry().getData());
212 if (!hasFollowers()) {
213 // for single node, the capture should happen after the apply state
214 // as we delete messages from the persistent journal which have made it to the snapshot
215 // capturing the snapshot before applying makes the persistent journal and snapshot out of sync
216 // and recovery shows data missing
217 context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry());
219 context.getSnapshotManager().trimLog(context.getLastApplied(), currentBehavior);
222 } else if (message instanceof ApplyJournalEntries){
223 ApplyJournalEntries applyEntries = (ApplyJournalEntries) message;
224 if(LOG.isDebugEnabled()) {
225 LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), applyEntries.getToIndex());
228 persistence().persist(applyEntries, NoopProcedure.instance());
230 } else if (message instanceof FindLeader) {
232 new FindLeaderReply(getLeaderAddress()),
235 } else if(message instanceof GetOnDemandRaftState) {
236 onGetOnDemandRaftStats();
237 } else if(!snapshotSupport.handleSnapshotMessage(message)) {
238 reusableBehaviorStateHolder.init(getCurrentBehavior());
240 setCurrentBehavior(currentBehavior.handleMessage(getSender(), message));
242 handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior());
246 protected RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() {
247 return new RaftActorSnapshotMessageSupport(context, currentBehavior,
248 getRaftActorSnapshotCohort());
251 private void onGetOnDemandRaftStats() {
252 // Debugging message to retrieve raft stats.
254 OnDemandRaftState.Builder builder = OnDemandRaftState.builder()
255 .commitIndex(context.getCommitIndex())
256 .currentTerm(context.getTermInformation().getCurrentTerm())
257 .inMemoryJournalDataSize(replicatedLog().dataSize())
258 .inMemoryJournalLogSize(replicatedLog().size())
259 .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing())
260 .lastApplied(context.getLastApplied())
261 .lastIndex(replicatedLog().lastIndex())
262 .lastTerm(replicatedLog().lastTerm())
263 .leader(getLeaderId())
264 .raftState(currentBehavior.state().toString())
265 .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex())
266 .snapshotIndex(replicatedLog().getSnapshotIndex())
267 .snapshotTerm(replicatedLog().getSnapshotTerm())
268 .votedFor(context.getTermInformation().getVotedFor())
269 .peerAddresses(new HashMap<>(context.getPeerAddresses()));
271 ReplicatedLogEntry lastLogEntry = getLastLogEntry();
272 if (lastLogEntry != null) {
273 builder.lastLogIndex(lastLogEntry.getIndex());
274 builder.lastLogTerm(lastLogEntry.getTerm());
277 if(getCurrentBehavior() instanceof AbstractLeader) {
278 AbstractLeader leader = (AbstractLeader)getCurrentBehavior();
279 Collection<String> followerIds = leader.getFollowerIds();
280 List<FollowerInfo> followerInfoList = Lists.newArrayListWithCapacity(followerIds.size());
281 for(String id: followerIds) {
282 final FollowerLogInformation info = leader.getFollower(id);
283 followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(),
284 info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity())));
287 builder.followerInfoList(followerInfoList);
290 sender().tell(builder.build(), self());
294 private void handleBehaviorChange(BehaviorStateHolder oldBehaviorState, RaftActorBehavior currentBehavior) {
295 RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior();
297 if (oldBehavior != currentBehavior){
301 String oldBehaviorLeaderId = oldBehavior == null ? null : oldBehaviorState.getLeaderId();
302 String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name();
304 // it can happen that the state has not changed but the leader has changed.
305 Optional<ActorRef> roleChangeNotifier = getRoleChangeNotifier();
306 if(!Objects.equal(oldBehaviorLeaderId, currentBehavior.getLeaderId())) {
307 if(roleChangeNotifier.isPresent()) {
308 roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId()), getSelf());
311 onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
314 if (roleChangeNotifier.isPresent() &&
315 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
316 roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName ,
317 currentBehavior.state().name()), getSelf());
321 protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId) {
322 return new LeaderStateChanged(memberId, leaderId);
326 public long snapshotSequenceNr() {
327 // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal,
328 // so that we can delete the persistent journal based on the saved sequence-number
329 // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot
330 // was saved and not the number we saved.
331 // We would want to override it , by asking akka to use the last-sequence number known to us.
332 return context.getSnapshotManager().getLastSequenceNumber();
336 * When a derived RaftActor needs to persist something it must call
343 protected void persistData(final ActorRef clientActor, final String identifier,
344 final Payload data) {
346 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
347 context.getReplicatedLog().lastIndex() + 1,
348 context.getTermInformation().getCurrentTerm(), data);
350 if(LOG.isDebugEnabled()) {
351 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
354 final RaftActorContext raftContext = getRaftActorContext();
356 replicatedLog().appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
358 public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
359 if (!hasFollowers()){
360 // Increment the Commit Index and the Last Applied values
361 raftContext.setCommitIndex(replicatedLogEntry.getIndex());
362 raftContext.setLastApplied(replicatedLogEntry.getIndex());
364 // Apply the state immediately.
365 self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry), self());
367 // Send a ApplyJournalEntries message so that we write the fact that we applied
368 // the state to durable storage
369 self().tell(new ApplyJournalEntries(replicatedLogEntry.getIndex()), self());
371 } else if (clientActor != null) {
372 context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry);
374 // Send message for replication
375 currentBehavior.handleMessage(getSelf(),
376 new Replicate(clientActor, identifier, replicatedLogEntry));
382 private ReplicatedLog replicatedLog() {
383 return context.getReplicatedLog();
386 protected String getId() {
387 return context.getId();
391 void setCurrentBehavior(RaftActorBehavior behavior) {
392 currentBehavior.setDelegate(behavior);
395 protected RaftActorBehavior getCurrentBehavior() {
396 return currentBehavior.getDelegate();
400 * Derived actors can call the isLeader method to check if the current
401 * RaftActor is the Leader or not
403 * @return true it this RaftActor is a Leader false otherwise
405 protected boolean isLeader() {
406 return context.getId().equals(currentBehavior.getLeaderId());
410 * Derived actor can call getLeader if they need a reference to the Leader.
411 * This would be useful for example in forwarding a request to an actor
412 * which is the leader
414 * @return A reference to the leader if known, null otherwise
416 protected ActorSelection getLeader(){
417 String leaderAddress = getLeaderAddress();
419 if(leaderAddress == null){
423 return context.actorSelection(leaderAddress);
428 * @return the current leader's id
430 protected String getLeaderId(){
431 return currentBehavior.getLeaderId();
434 protected RaftState getRaftState() {
435 return currentBehavior.state();
438 protected ReplicatedLogEntry getLastLogEntry() {
439 return replicatedLog().last();
442 protected Long getCurrentTerm(){
443 return context.getTermInformation().getCurrentTerm();
446 protected Long getCommitIndex(){
447 return context.getCommitIndex();
450 protected Long getLastApplied(){
451 return context.getLastApplied();
454 protected RaftActorContext getRaftActorContext() {
458 protected void updateConfigParams(ConfigParams configParams) {
459 context.setConfigParams(configParams);
462 public final DataPersistenceProvider persistence() {
463 return delegatingPersistenceProvider.getDelegate();
466 public void setPersistence(DataPersistenceProvider provider) {
467 delegatingPersistenceProvider.setDelegate(provider);
470 protected void setPersistence(boolean persistent) {
472 setPersistence(new PersistentDataProvider(this));
474 setPersistence(new NonPersistentDataProvider() {
476 * The way snapshotting works is,
478 * <li> RaftActor calls createSnapshot on the Shard
479 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
480 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save
481 * the snapshot. The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the
482 * RaftActor gets SaveSnapshot success it commits the snapshot to the in-memory journal. This
483 * commitSnapshot is mimicking what is done in SaveSnapshotSuccess.
487 public void saveSnapshot(Object o) {
488 // Make saving Snapshot successful
489 // Committing the snapshot here would end up calling commit in the creating state which would
490 // be a state violation. That's why now we send a message to commit the snapshot.
491 self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self());
498 * setPeerAddress sets the address of a known peer at a later time.
500 * This is to account for situations where a we know that a peer
501 * exists but we do not know an address up-front. This may also be used in
502 * situations where a known peer starts off in a different location and we
503 * need to change it's address
505 * Note that if the peerId does not match the list of peers passed to
506 * this actor during construction an IllegalStateException will be thrown.
511 protected void setPeerAddress(String peerId, String peerAddress){
512 context.setPeerAddress(peerId, peerAddress);
516 * The applyState method will be called by the RaftActor when some data
517 * needs to be applied to the actor's state
519 * @param clientActor A reference to the client who sent this message. This
520 * is the same reference that was passed to persistData
521 * by the derived actor. clientActor may be null when
522 * the RaftActor is behaving as a follower or during
524 * @param identifier The identifier of the persisted data. This is also
525 * the same identifier that was passed to persistData by
526 * the derived actor. identifier may be null when
527 * the RaftActor is behaving as a follower or during
529 * @param data A piece of data that was persisted by the persistData call.
530 * This should NEVER be null.
532 protected abstract void applyState(ActorRef clientActor, String identifier,
536 * Returns the RaftActorRecoveryCohort to participate in persistence recovery.
539 protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort();
542 * This method is called when recovery is complete.
544 protected abstract void onRecoveryComplete();
547 * Returns the RaftActorSnapshotCohort to participate in persistence recovery.
550 protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort();
553 * This method will be called by the RaftActor when the state of the
554 * RaftActor changes. The derived actor can then use methods like
555 * isLeader or getLeader to do something useful
557 protected abstract void onStateChanged();
560 * Notifier Actor for this RaftActor to notify when a role change happens
561 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
563 protected abstract Optional<ActorRef> getRoleChangeNotifier();
565 protected void onLeaderChanged(String oldLeader, String newLeader){};
567 private String getLeaderAddress(){
569 return getSelf().path().toString();
571 String leaderId = currentBehavior.getLeaderId();
572 if (leaderId == null) {
575 String peerAddress = context.getPeerAddress(leaderId);
576 if(LOG.isDebugEnabled()) {
577 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
578 persistenceId(), leaderId, peerAddress);
584 protected boolean hasFollowers(){
585 return getRaftActorContext().hasFollowers();
589 * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries}
590 * whose type for fromIndex is long instead of int. This class was kept for backwards
591 * compatibility with Helium.
593 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
594 @SuppressWarnings("serial")
596 static class DeleteEntries implements Serializable {
597 private final int fromIndex;
599 public DeleteEntries(int fromIndex) {
600 this.fromIndex = fromIndex;
603 public int getFromIndex() {
609 * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm}
610 * which has serialVersionUID set. This class was kept for backwards compatibility with Helium.
612 // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility.
613 @SuppressWarnings("serial")
615 static class UpdateElectionTerm implements Serializable {
616 private final long currentTerm;
617 private final String votedFor;
619 public UpdateElectionTerm(long currentTerm, String votedFor) {
620 this.currentTerm = currentTerm;
621 this.votedFor = votedFor;
624 public long getCurrentTerm() {
628 public String getVotedFor() {
633 private static class BehaviorStateHolder {
634 private RaftActorBehavior behavior;
635 private String leaderId;
637 void init(RaftActorBehavior behavior) {
638 this.behavior = behavior;
639 this.leaderId = behavior != null ? behavior.getLeaderId() : null;
642 RaftActorBehavior getBehavior() {
646 String getLeaderId() {