2 * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
9 package org.opendaylight.controller.cluster.raft;
11 import akka.actor.ActorRef;
12 import akka.actor.ActorSelection;
13 import akka.event.Logging;
14 import akka.event.LoggingAdapter;
15 import akka.japi.Procedure;
16 import akka.persistence.RecoveryCompleted;
17 import akka.persistence.SaveSnapshotFailure;
18 import akka.persistence.SaveSnapshotSuccess;
19 import akka.persistence.SnapshotOffer;
20 import akka.persistence.SnapshotSelectionCriteria;
21 import akka.persistence.UntypedPersistentActor;
22 import com.google.common.base.Optional;
23 import com.google.protobuf.ByteString;
24 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
25 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
26 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
27 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
28 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
29 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
30 import org.opendaylight.controller.cluster.raft.behaviors.Candidate;
31 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
32 import org.opendaylight.controller.cluster.raft.behaviors.Leader;
33 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
34 import org.opendaylight.controller.cluster.raft.client.messages.AddRaftPeer;
35 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
36 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
37 import org.opendaylight.controller.cluster.raft.client.messages.RemoveRaftPeer;
38 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
39 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
40 import org.opendaylight.controller.protobuff.messages.cluster.raft.AppendEntriesMessages;
42 import java.io.Serializable;
46 * RaftActor encapsulates a state machine that needs to be kept synchronized
47 * in a cluster. It implements the RAFT algorithm as described in the paper
48 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
49 * In Search of an Understandable Consensus Algorithm</a>
51 * RaftActor has 3 states and each state has a certain behavior associated
52 * with it. A Raft actor can behave as,
55 * <li> A Follower (or) </li>
56 * <li> A Candidate </li>
60 * A RaftActor MUST be a Leader in order to accept requests from clients to
61 * change the state of it's encapsulated state machine. Once a RaftActor becomes
62 * a Leader it is also responsible for ensuring that all followers ultimately
63 * have the same log and therefore the same state machine as itself.
66 * The current behavior of a RaftActor determines how election for leadership
67 * is initiated and how peer RaftActors react to request for votes.
70 * Each RaftActor also needs to know the current election term. It uses this
71 * information for a couple of things. One is to simply figure out who it
72 * voted for in the last election. Another is to figure out if the message
73 * it received to update it's state is stale.
76 * The RaftActor uses akka-persistence to store it's replicated log.
77 * Furthermore through it's behaviors a Raft Actor determines
80 * <li> when a log entry should be persisted </li>
81 * <li> when a log entry should be applied to the state machine (and) </li>
82 * <li> when a snapshot should be saved </li>
85 public abstract class RaftActor extends UntypedPersistentActor {
86 protected final LoggingAdapter LOG =
87 Logging.getLogger(getContext().system(), this);
90 * The current state determines the current behavior of a RaftActor
91 * A Raft Actor always starts off in the Follower State
93 private RaftActorBehavior currentBehavior;
96 * This context should NOT be passed directly to any other actor it is
97 * only to be consumed by the RaftActorBehaviors
99 private RaftActorContext context;
102 * The in-memory journal
104 private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
106 private CaptureSnapshot captureSnapshot = null;
108 private volatile boolean hasSnapshotCaptureInitiated = false;
110 public RaftActor(String id, Map<String, String> peerAddresses) {
111 this(id, peerAddresses, Optional.<ConfigParams>absent());
114 public RaftActor(String id, Map<String, String> peerAddresses,
115 Optional<ConfigParams> configParams) {
117 context = new RaftActorContextImpl(this.getSelf(),
118 this.getContext(), id, new ElectionTermImpl(),
119 -1, -1, replicatedLog, peerAddresses,
120 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
124 @Override public void onReceiveRecover(Object message) {
125 if (message instanceof SnapshotOffer) {
126 LOG.info("SnapshotOffer called..");
127 SnapshotOffer offer = (SnapshotOffer) message;
128 Snapshot snapshot = (Snapshot) offer.snapshot();
130 // Create a replicated log with the snapshot information
131 // The replicated log can be used later on to retrieve this snapshot
132 // when we need to install it on a peer
133 replicatedLog = new ReplicatedLogImpl(snapshot);
135 context.setReplicatedLog(replicatedLog);
136 context.setLastApplied(snapshot.getLastAppliedIndex());
138 LOG.info("Applied snapshot to replicatedLog. " +
139 "snapshotIndex={}, snapshotTerm={}, journal-size={}",
140 replicatedLog.snapshotIndex, replicatedLog.snapshotTerm,
144 // Apply the snapshot to the actors state
145 applySnapshot(ByteString.copyFrom(snapshot.getState()));
147 } else if (message instanceof ReplicatedLogEntry) {
148 ReplicatedLogEntry logEntry = (ReplicatedLogEntry) message;
150 // Apply State immediately
151 replicatedLog.append(logEntry);
152 applyState(null, "recovery", logEntry.getData());
153 context.setLastApplied(logEntry.getIndex());
154 context.setCommitIndex(logEntry.getIndex());
155 } else if (message instanceof DeleteEntries) {
156 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
157 } else if (message instanceof UpdateElectionTerm) {
158 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor());
159 } else if (message instanceof RecoveryCompleted) {
160 if(LOG.isDebugEnabled()) {
162 "RecoveryCompleted - Switching actor to Follower - " +
163 "Persistence Id = " + persistenceId() +
164 " Last index in log:{}, snapshotIndex={}, snapshotTerm={}, " +
166 replicatedLog.lastIndex(), replicatedLog.snapshotIndex,
167 replicatedLog.snapshotTerm, replicatedLog.size()
170 currentBehavior = switchBehavior(RaftState.Follower);
175 @Override public void onReceiveCommand(Object message) {
176 if (message instanceof ApplyState){
177 ApplyState applyState = (ApplyState) message;
179 if(LOG.isDebugEnabled()) {
180 LOG.debug("Applying state for log index {} data {}",
181 applyState.getReplicatedLogEntry().getIndex(),
182 applyState.getReplicatedLogEntry().getData());
185 applyState(applyState.getClientActor(), applyState.getIdentifier(),
186 applyState.getReplicatedLogEntry().getData());
188 } else if(message instanceof ApplySnapshot ) {
189 Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
191 if(LOG.isDebugEnabled()) {
192 LOG.debug("ApplySnapshot called on Follower Actor " +
193 "snapshotIndex:{}, snapshotTerm:{}", snapshot.getLastAppliedIndex(),
194 snapshot.getLastAppliedTerm()
197 applySnapshot(ByteString.copyFrom(snapshot.getState()));
199 //clears the followers log, sets the snapshot index to ensure adjusted-index works
200 replicatedLog = new ReplicatedLogImpl(snapshot);
201 context.setReplicatedLog(replicatedLog);
202 context.setLastApplied(snapshot.getLastAppliedIndex());
204 } else if (message instanceof FindLeader) {
206 new FindLeaderReply(getLeaderAddress()),
210 } else if (message instanceof SaveSnapshotSuccess) {
211 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
212 LOG.info("SaveSnapshotSuccess received for snapshot");
214 context.getReplicatedLog().snapshotCommit();
216 // TODO: Not sure if we want to be this aggressive with trimming stuff
217 trimPersistentData(success.metadata().sequenceNr());
219 } else if (message instanceof SaveSnapshotFailure) {
220 SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
222 LOG.info("saveSnapshotFailure.metadata():{}", saveSnapshotFailure.metadata().toString());
223 LOG.error(saveSnapshotFailure.cause(), "SaveSnapshotFailure received for snapshot Cause:");
225 context.getReplicatedLog().snapshotRollback();
227 LOG.info("Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
228 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}",
229 context.getReplicatedLog().getSnapshotIndex(),
230 context.getReplicatedLog().getSnapshotTerm(),
231 context.getReplicatedLog().size());
233 } else if (message instanceof AddRaftPeer){
235 // FIXME : Do not add raft peers like this.
236 // When adding a new Peer we have to ensure that the a majority of
237 // the peers know about the new Peer. Doing it this way may cause
238 // a situation where multiple Leaders may emerge
239 AddRaftPeer arp = (AddRaftPeer)message;
240 context.addToPeers(arp.getName(), arp.getAddress());
242 } else if (message instanceof RemoveRaftPeer){
244 RemoveRaftPeer rrp = (RemoveRaftPeer)message;
245 context.removePeer(rrp.getName());
247 } else if (message instanceof CaptureSnapshot) {
248 LOG.info("CaptureSnapshot received by actor");
249 CaptureSnapshot cs = (CaptureSnapshot)message;
250 captureSnapshot = cs;
253 } else if (message instanceof CaptureSnapshotReply){
254 LOG.info("CaptureSnapshotReply received by actor");
255 CaptureSnapshotReply csr = (CaptureSnapshotReply) message;
257 ByteString stateInBytes = csr.getSnapshot();
258 LOG.info("CaptureSnapshotReply stateInBytes size:{}", stateInBytes.size());
259 handleCaptureSnapshotReply(stateInBytes);
262 if (!(message instanceof AppendEntriesMessages.AppendEntries)
263 && !(message instanceof AppendEntriesReply) && !(message instanceof SendHeartBeat)) {
264 if(LOG.isDebugEnabled()) {
265 LOG.debug("onReceiveCommand: message:" + message.getClass());
270 currentBehavior.handleMessage(getSender(), message);
271 RaftActorBehavior oldBehavior = currentBehavior;
272 currentBehavior = switchBehavior(state);
273 if(oldBehavior != currentBehavior){
277 onLeaderChanged(oldBehavior.getLeaderId(), currentBehavior.getLeaderId());
281 public java.util.Set<String> getPeers() {
282 return context.getPeerAddresses().keySet();
285 protected String getReplicatedLogState() {
286 return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex()
287 + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm()
288 + ", im-mem journal size=" + context.getReplicatedLog().size();
293 * When a derived RaftActor needs to persist something it must call
300 protected void persistData(ActorRef clientActor, String identifier,
303 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
304 context.getReplicatedLog().lastIndex() + 1,
305 context.getTermInformation().getCurrentTerm(), data);
307 if(LOG.isDebugEnabled()) {
308 LOG.debug("Persist data {}", replicatedLogEntry);
312 .appendAndPersist(clientActor, identifier, replicatedLogEntry);
315 protected String getId() {
316 return context.getId();
320 * Derived actors can call the isLeader method to check if the current
321 * RaftActor is the Leader or not
323 * @return true it this RaftActor is a Leader false otherwise
325 protected boolean isLeader() {
326 return context.getId().equals(currentBehavior.getLeaderId());
330 * Derived actor can call getLeader if they need a reference to the Leader.
331 * This would be useful for example in forwarding a request to an actor
332 * which is the leader
334 * @return A reference to the leader if known, null otherwise
336 protected ActorSelection getLeader(){
337 String leaderAddress = getLeaderAddress();
339 if(leaderAddress == null){
343 return context.actorSelection(leaderAddress);
348 * @return the current leader's id
350 protected String getLeaderId(){
351 return currentBehavior.getLeaderId();
354 protected RaftState getRaftState() {
355 return currentBehavior.state();
358 protected ReplicatedLogEntry getLastLogEntry() {
359 return replicatedLog.last();
362 protected Long getCurrentTerm(){
363 return context.getTermInformation().getCurrentTerm();
366 protected Long getCommitIndex(){
367 return context.getCommitIndex();
370 protected Long getLastApplied(){
371 return context.getLastApplied();
375 * setPeerAddress sets the address of a known peer at a later time.
377 * This is to account for situations where a we know that a peer
378 * exists but we do not know an address up-front. This may also be used in
379 * situations where a known peer starts off in a different location and we
380 * need to change it's address
382 * Note that if the peerId does not match the list of peers passed to
383 * this actor during construction an IllegalStateException will be thrown.
388 protected void setPeerAddress(String peerId, String peerAddress){
389 context.setPeerAddress(peerId, peerAddress);
395 * The applyState method will be called by the RaftActor when some data
396 * needs to be applied to the actor's state
398 * @param clientActor A reference to the client who sent this message. This
399 * is the same reference that was passed to persistData
400 * by the derived actor. clientActor may be null when
401 * the RaftActor is behaving as a follower or during
403 * @param identifier The identifier of the persisted data. This is also
404 * the same identifier that was passed to persistData by
405 * the derived actor. identifier may be null when
406 * the RaftActor is behaving as a follower or during
408 * @param data A piece of data that was persisted by the persistData call.
409 * This should NEVER be null.
411 protected abstract void applyState(ActorRef clientActor, String identifier,
415 * This method will be called by the RaftActor when a snapshot needs to be
416 * created. The derived actor should respond with its current state.
418 * During recovery the state that is returned by the derived actor will
419 * be passed back to it by calling the applySnapshot method
421 * @return The current state of the actor
423 protected abstract void createSnapshot();
426 * This method will be called by the RaftActor during recovery to
427 * reconstruct the state of the actor.
429 * This method may also be called at any other point during normal
430 * operations when the derived actor is out of sync with it's peers
431 * and the only way to bring it in sync is by applying a snapshot
433 * @param snapshot A snapshot of the state of the actor
435 protected abstract void applySnapshot(ByteString snapshot);
438 * This method will be called by the RaftActor when the state of the
439 * RaftActor changes. The derived actor can then use methods like
440 * isLeader or getLeader to do something useful
442 protected abstract void onStateChanged();
444 protected void onLeaderChanged(String oldLeader, String newLeader){};
446 private RaftActorBehavior switchBehavior(RaftState state) {
447 if (currentBehavior != null) {
448 if (currentBehavior.state() == state) {
449 return currentBehavior;
451 LOG.info("Switching from state " + currentBehavior.state() + " to "
455 currentBehavior.close();
456 } catch (Exception e) {
458 "Failed to close behavior : " + currentBehavior.state());
462 LOG.info("Switching behavior to " + state);
464 RaftActorBehavior behavior = null;
465 if (state == RaftState.Candidate) {
466 behavior = new Candidate(context);
467 } else if (state == RaftState.Follower) {
468 behavior = new Follower(context);
470 behavior = new Leader(context);
478 private void trimPersistentData(long sequenceNumber) {
479 // Trim akka snapshots
480 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
481 // For now guessing that it is ANDed.
482 deleteSnapshots(new SnapshotSelectionCriteria(
483 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
486 deleteMessages(sequenceNumber);
489 private String getLeaderAddress(){
491 return getSelf().path().toString();
493 String leaderId = currentBehavior.getLeaderId();
494 if (leaderId == null) {
497 String peerAddress = context.getPeerAddress(leaderId);
498 if(LOG.isDebugEnabled()) {
499 LOG.debug("getLeaderAddress leaderId = " + leaderId + " peerAddress = "
506 private void handleCaptureSnapshotReply(ByteString stateInBytes) {
507 // create a snapshot object from the state provided and save it
508 // when snapshot is saved async, SaveSnapshotSuccess is raised.
510 Snapshot sn = Snapshot.create(stateInBytes.toByteArray(),
511 context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
512 captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
513 captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
517 LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
519 //be greedy and remove entries from in-mem journal which are in the snapshot
520 // and update snapshotIndex and snapshotTerm without waiting for the success,
522 context.getReplicatedLog().snapshotPreCommit(stateInBytes,
523 captureSnapshot.getLastAppliedIndex(),
524 captureSnapshot.getLastAppliedTerm());
526 LOG.info("Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
527 "and term:{}", captureSnapshot.getLastAppliedIndex(),
528 captureSnapshot.getLastAppliedTerm());
530 captureSnapshot = null;
531 hasSnapshotCaptureInitiated = false;
535 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
537 public ReplicatedLogImpl(Snapshot snapshot) {
538 super(ByteString.copyFrom(snapshot.getState()),
539 snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
540 snapshot.getUnAppliedEntries());
543 public ReplicatedLogImpl() {
547 @Override public void removeFromAndPersist(long logEntryIndex) {
548 int adjustedIndex = adjustedIndex(logEntryIndex);
550 if (adjustedIndex < 0) {
554 // FIXME: Maybe this should be done after the command is saved
555 journal.subList(adjustedIndex , journal.size()).clear();
557 persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
559 @Override public void apply(DeleteEntries param)
561 //FIXME : Doing nothing for now
566 @Override public void appendAndPersist(
567 final ReplicatedLogEntry replicatedLogEntry) {
568 appendAndPersist(null, null, replicatedLogEntry);
571 public void appendAndPersist(final ActorRef clientActor,
572 final String identifier,
573 final ReplicatedLogEntry replicatedLogEntry) {
574 context.getLogger().debug(
575 "Append log entry and persist {} ", replicatedLogEntry);
576 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
577 journal.add(replicatedLogEntry);
579 // When persisting events with persist it is guaranteed that the
580 // persistent actor will not receive further commands between the
581 // persist call and the execution(s) of the associated event
582 // handler. This also holds for multiple persist calls in context
583 // of a single command.
584 persist(replicatedLogEntry,
585 new Procedure<ReplicatedLogEntry>() {
586 public void apply(ReplicatedLogEntry evt) throws Exception {
587 // when a snaphsot is being taken, captureSnapshot != null
588 if (hasSnapshotCaptureInitiated == false &&
589 journal.size() % context.getConfigParams().getSnapshotBatchCount() == 0) {
591 LOG.info("Initiating Snapshot Capture..");
592 long lastAppliedIndex = -1;
593 long lastAppliedTerm = -1;
595 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
596 if (lastAppliedEntry != null) {
597 lastAppliedIndex = lastAppliedEntry.getIndex();
598 lastAppliedTerm = lastAppliedEntry.getTerm();
601 if(LOG.isDebugEnabled()) {
602 LOG.debug("Snapshot Capture logSize: {}", journal.size());
603 LOG.debug("Snapshot Capture lastApplied:{} ",
604 context.getLastApplied());
605 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
606 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
609 // send a CaptureSnapshot to self to make the expensive operation async.
610 getSelf().tell(new CaptureSnapshot(
611 lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm),
613 hasSnapshotCaptureInitiated = true;
615 // Send message for replication
616 if (clientActor != null) {
617 currentBehavior.handleMessage(getSelf(),
618 new Replicate(clientActor, identifier,
629 private static class DeleteEntries implements Serializable {
630 private final int fromIndex;
633 public DeleteEntries(int fromIndex) {
634 this.fromIndex = fromIndex;
637 public int getFromIndex() {
643 private class ElectionTermImpl implements ElectionTerm {
645 * Identifier of the actor whose election term information this is
647 private long currentTerm = 0;
648 private String votedFor = null;
650 public long getCurrentTerm() {
654 public String getVotedFor() {
658 @Override public void update(long currentTerm, String votedFor) {
659 if(LOG.isDebugEnabled()) {
660 LOG.debug("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
662 this.currentTerm = currentTerm;
663 this.votedFor = votedFor;
667 public void updateAndPersist(long currentTerm, String votedFor){
668 update(currentTerm, votedFor);
669 // FIXME : Maybe first persist then update the state
670 persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
672 @Override public void apply(UpdateElectionTerm param)
680 private static class UpdateElectionTerm implements Serializable {
681 private final long currentTerm;
682 private final String votedFor;
684 public UpdateElectionTerm(long currentTerm, String votedFor) {
685 this.currentTerm = currentTerm;
686 this.votedFor = votedFor;
689 public long getCurrentTerm() {
693 public String getVotedFor() {