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.debug("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.debug("Applied snapshot to replicatedLog. " +
139 "snapshotIndex={}, snapshotTerm={}, journal-size={}",
140 replicatedLog.snapshotIndex, replicatedLog.snapshotTerm,
141 replicatedLog.size());
143 // Apply the snapshot to the actors state
144 applySnapshot(ByteString.copyFrom(snapshot.getState()));
146 } else if (message instanceof ReplicatedLogEntry) {
147 replicatedLog.append((ReplicatedLogEntry) message);
148 } else if (message instanceof DeleteEntries) {
149 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
150 } else if (message instanceof UpdateElectionTerm) {
151 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor());
152 } else if (message instanceof RecoveryCompleted) {
154 "RecoveryCompleted - Switching actor to Follower - " +
155 "Last index in log:{}, snapshotIndex={}, snapshotTerm={}, " +
157 replicatedLog.lastIndex(), replicatedLog.snapshotIndex,
158 replicatedLog.snapshotTerm, replicatedLog.size());
159 currentBehavior = switchBehavior(RaftState.Follower);
164 @Override public void onReceiveCommand(Object message) {
165 if (message instanceof ApplyState){
166 ApplyState applyState = (ApplyState) message;
168 LOG.debug("Applying state for log index {} data {}",
169 applyState.getReplicatedLogEntry().getIndex(),
170 applyState.getReplicatedLogEntry().getData());
172 applyState(applyState.getClientActor(), applyState.getIdentifier(),
173 applyState.getReplicatedLogEntry().getData());
175 } else if(message instanceof ApplySnapshot ) {
176 Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
178 LOG.debug("ApplySnapshot called on Follower Actor " +
179 "snapshotIndex:{}, snapshotTerm:{}", snapshot.getLastAppliedIndex(),
180 snapshot.getLastAppliedTerm());
181 applySnapshot(ByteString.copyFrom(snapshot.getState()));
183 //clears the followers log, sets the snapshot index to ensure adjusted-index works
184 replicatedLog = new ReplicatedLogImpl(snapshot);
185 context.setReplicatedLog(replicatedLog);
186 context.setLastApplied(snapshot.getLastAppliedIndex());
188 } else if (message instanceof FindLeader) {
190 new FindLeaderReply(getLeaderAddress()),
194 } else if (message instanceof SaveSnapshotSuccess) {
195 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
196 LOG.info("SaveSnapshotSuccess received for snapshot");
198 context.getReplicatedLog().snapshotCommit();
200 // TODO: Not sure if we want to be this aggressive with trimming stuff
201 trimPersistentData(success.metadata().sequenceNr());
203 } else if (message instanceof SaveSnapshotFailure) {
204 SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
206 LOG.info("saveSnapshotFailure.metadata():{}", saveSnapshotFailure.metadata().toString());
207 LOG.error(saveSnapshotFailure.cause(), "SaveSnapshotFailure received for snapshot Cause:");
209 context.getReplicatedLog().snapshotRollback();
211 LOG.info("Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
212 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}",
213 context.getReplicatedLog().getSnapshotIndex(),
214 context.getReplicatedLog().getSnapshotTerm(),
215 context.getReplicatedLog().size());
217 } else if (message instanceof AddRaftPeer){
219 // FIXME : Do not add raft peers like this.
220 // When adding a new Peer we have to ensure that the a majority of
221 // the peers know about the new Peer. Doing it this way may cause
222 // a situation where multiple Leaders may emerge
223 AddRaftPeer arp = (AddRaftPeer)message;
224 context.addToPeers(arp.getName(), arp.getAddress());
226 } else if (message instanceof RemoveRaftPeer){
228 RemoveRaftPeer rrp = (RemoveRaftPeer)message;
229 context.removePeer(rrp.getName());
231 } else if (message instanceof CaptureSnapshot) {
232 LOG.debug("CaptureSnapshot received by actor");
233 CaptureSnapshot cs = (CaptureSnapshot)message;
234 captureSnapshot = cs;
237 } else if (message instanceof CaptureSnapshotReply){
238 LOG.debug("CaptureSnapshotReply received by actor");
239 CaptureSnapshotReply csr = (CaptureSnapshotReply) message;
241 ByteString stateInBytes = csr.getSnapshot();
242 LOG.debug("CaptureSnapshotReply stateInBytes size:{}", stateInBytes.size());
243 handleCaptureSnapshotReply(stateInBytes);
246 if (!(message instanceof AppendEntriesMessages.AppendEntries)
247 && !(message instanceof AppendEntriesReply) && !(message instanceof SendHeartBeat)) {
248 LOG.debug("onReceiveCommand: message:" + message.getClass());
252 currentBehavior.handleMessage(getSender(), message);
253 RaftActorBehavior oldBehavior = currentBehavior;
254 currentBehavior = switchBehavior(state);
255 if(oldBehavior != currentBehavior){
261 public java.util.Set<String> getPeers() {
262 return context.getPeerAddresses().keySet();
265 protected String getReplicatedLogState() {
266 return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex()
267 + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm()
268 + ", im-mem journal size=" + context.getReplicatedLog().size();
273 * When a derived RaftActor needs to persist something it must call
280 protected void persistData(ActorRef clientActor, String identifier,
283 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
284 context.getReplicatedLog().lastIndex() + 1,
285 context.getTermInformation().getCurrentTerm(), data);
287 LOG.debug("Persist data {}", replicatedLogEntry);
290 .appendAndPersist(clientActor, identifier, replicatedLogEntry);
293 protected String getId() {
294 return context.getId();
298 * Derived actors can call the isLeader method to check if the current
299 * RaftActor is the Leader or not
301 * @return true it this RaftActor is a Leader false otherwise
303 protected boolean isLeader() {
304 return context.getId().equals(currentBehavior.getLeaderId());
308 * Derived actor can call getLeader if they need a reference to the Leader.
309 * This would be useful for example in forwarding a request to an actor
310 * which is the leader
312 * @return A reference to the leader if known, null otherwise
314 protected ActorSelection getLeader(){
315 String leaderAddress = getLeaderAddress();
317 if(leaderAddress == null){
321 return context.actorSelection(leaderAddress);
326 * @return the current leader's id
328 protected String getLeaderId(){
329 return currentBehavior.getLeaderId();
332 protected RaftState getRaftState() {
333 return currentBehavior.state();
336 protected ReplicatedLogEntry getLastLogEntry() {
337 return replicatedLog.last();
340 protected Long getCurrentTerm(){
341 return context.getTermInformation().getCurrentTerm();
344 protected Long getCommitIndex(){
345 return context.getCommitIndex();
348 protected Long getLastApplied(){
349 return context.getLastApplied();
353 * setPeerAddress sets the address of a known peer at a later time.
355 * This is to account for situations where a we know that a peer
356 * exists but we do not know an address up-front. This may also be used in
357 * situations where a known peer starts off in a different location and we
358 * need to change it's address
360 * Note that if the peerId does not match the list of peers passed to
361 * this actor during construction an IllegalStateException will be thrown.
366 protected void setPeerAddress(String peerId, String peerAddress){
367 context.setPeerAddress(peerId, peerAddress);
373 * The applyState method will be called by the RaftActor when some data
374 * needs to be applied to the actor's state
376 * @param clientActor A reference to the client who sent this message. This
377 * is the same reference that was passed to persistData
378 * by the derived actor. clientActor may be null when
379 * the RaftActor is behaving as a follower or during
381 * @param identifier The identifier of the persisted data. This is also
382 * the same identifier that was passed to persistData by
383 * the derived actor. identifier may be null when
384 * the RaftActor is behaving as a follower or during
386 * @param data A piece of data that was persisted by the persistData call.
387 * This should NEVER be null.
389 protected abstract void applyState(ActorRef clientActor, String identifier,
393 * This method will be called by the RaftActor when a snapshot needs to be
394 * created. The derived actor should respond with its current state.
396 * During recovery the state that is returned by the derived actor will
397 * be passed back to it by calling the applySnapshot method
399 * @return The current state of the actor
401 protected abstract void createSnapshot();
404 * This method will be called by the RaftActor during recovery to
405 * reconstruct the state of the actor.
407 * This method may also be called at any other point during normal
408 * operations when the derived actor is out of sync with it's peers
409 * and the only way to bring it in sync is by applying a snapshot
411 * @param snapshot A snapshot of the state of the actor
413 protected abstract void applySnapshot(ByteString snapshot);
416 * This method will be called by the RaftActor when the state of the
417 * RaftActor changes. The derived actor can then use methods like
418 * isLeader or getLeader to do something useful
420 protected abstract void onStateChanged();
422 private RaftActorBehavior switchBehavior(RaftState state) {
423 if (currentBehavior != null) {
424 if (currentBehavior.state() == state) {
425 return currentBehavior;
427 LOG.info("Switching from state " + currentBehavior.state() + " to "
431 currentBehavior.close();
432 } catch (Exception e) {
434 "Failed to close behavior : " + currentBehavior.state());
438 LOG.info("Switching behavior to " + state);
440 RaftActorBehavior behavior = null;
441 if (state == RaftState.Candidate) {
442 behavior = new Candidate(context);
443 } else if (state == RaftState.Follower) {
444 behavior = new Follower(context);
446 behavior = new Leader(context);
454 private void trimPersistentData(long sequenceNumber) {
455 // Trim akka snapshots
456 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
457 // For now guessing that it is ANDed.
458 deleteSnapshots(new SnapshotSelectionCriteria(
459 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
462 deleteMessages(sequenceNumber);
465 private String getLeaderAddress(){
467 return getSelf().path().toString();
469 String leaderId = currentBehavior.getLeaderId();
470 if (leaderId == null) {
473 String peerAddress = context.getPeerAddress(leaderId);
474 LOG.debug("getLeaderAddress leaderId = " + leaderId + " peerAddress = "
480 private void handleCaptureSnapshotReply(ByteString stateInBytes) {
481 // create a snapshot object from the state provided and save it
482 // when snapshot is saved async, SaveSnapshotSuccess is raised.
484 Snapshot sn = Snapshot.create(stateInBytes.toByteArray(),
485 context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
486 captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
487 captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
491 LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
493 //be greedy and remove entries from in-mem journal which are in the snapshot
494 // and update snapshotIndex and snapshotTerm without waiting for the success,
496 context.getReplicatedLog().snapshotPreCommit(stateInBytes,
497 captureSnapshot.getLastAppliedIndex(),
498 captureSnapshot.getLastAppliedTerm());
500 LOG.info("Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
501 "and term:{}", captureSnapshot.getLastAppliedIndex(),
502 captureSnapshot.getLastAppliedTerm());
504 captureSnapshot = null;
505 hasSnapshotCaptureInitiated = false;
509 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
511 public ReplicatedLogImpl(Snapshot snapshot) {
512 super(ByteString.copyFrom(snapshot.getState()),
513 snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
514 snapshot.getUnAppliedEntries());
517 public ReplicatedLogImpl() {
521 @Override public void removeFromAndPersist(long logEntryIndex) {
522 int adjustedIndex = adjustedIndex(logEntryIndex);
524 if (adjustedIndex < 0) {
528 // FIXME: Maybe this should be done after the command is saved
529 journal.subList(adjustedIndex , journal.size()).clear();
531 persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
533 @Override public void apply(DeleteEntries param)
535 //FIXME : Doing nothing for now
540 @Override public void appendAndPersist(
541 final ReplicatedLogEntry replicatedLogEntry) {
542 appendAndPersist(null, null, replicatedLogEntry);
545 public void appendAndPersist(final ActorRef clientActor,
546 final String identifier,
547 final ReplicatedLogEntry replicatedLogEntry) {
548 context.getLogger().debug(
549 "Append log entry and persist {} ", replicatedLogEntry);
550 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
551 journal.add(replicatedLogEntry);
553 // When persisting events with persist it is guaranteed that the
554 // persistent actor will not receive further commands between the
555 // persist call and the execution(s) of the associated event
556 // handler. This also holds for multiple persist calls in context
557 // of a single command.
558 persist(replicatedLogEntry,
559 new Procedure<ReplicatedLogEntry>() {
560 public void apply(ReplicatedLogEntry evt) throws Exception {
561 // when a snaphsot is being taken, captureSnapshot != null
562 if (hasSnapshotCaptureInitiated == false &&
563 journal.size() % context.getConfigParams().getSnapshotBatchCount() == 0) {
565 LOG.info("Initiating Snapshot Capture..");
566 long lastAppliedIndex = -1;
567 long lastAppliedTerm = -1;
569 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
570 if (lastAppliedEntry != null) {
571 lastAppliedIndex = lastAppliedEntry.getIndex();
572 lastAppliedTerm = lastAppliedEntry.getTerm();
575 LOG.debug("Snapshot Capture logSize: {}", journal.size());
576 LOG.debug("Snapshot Capture lastApplied:{} ", context.getLastApplied());
577 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
578 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
580 // send a CaptureSnapshot to self to make the expensive operation async.
581 getSelf().tell(new CaptureSnapshot(
582 lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm),
584 hasSnapshotCaptureInitiated = true;
586 // Send message for replication
587 if (clientActor != null) {
588 currentBehavior.handleMessage(getSelf(),
589 new Replicate(clientActor, identifier,
600 private static class DeleteEntries implements Serializable {
601 private final int fromIndex;
604 public DeleteEntries(int fromIndex) {
605 this.fromIndex = fromIndex;
608 public int getFromIndex() {
614 private class ElectionTermImpl implements ElectionTerm {
616 * Identifier of the actor whose election term information this is
618 private long currentTerm = 0;
619 private String votedFor = null;
621 public long getCurrentTerm() {
625 public String getVotedFor() {
629 @Override public void update(long currentTerm, String votedFor) {
630 LOG.debug("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
632 this.currentTerm = currentTerm;
633 this.votedFor = votedFor;
637 public void updateAndPersist(long currentTerm, String votedFor){
638 update(currentTerm, votedFor);
639 // FIXME : Maybe first persist then update the state
640 persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
642 @Override public void apply(UpdateElectionTerm param)
650 private static class UpdateElectionTerm implements Serializable {
651 private final long currentTerm;
652 private final String votedFor;
654 public UpdateElectionTerm(long currentTerm, String votedFor) {
655 this.currentTerm = currentTerm;
656 this.votedFor = votedFor;
659 public long getCurrentTerm() {
663 public String getVotedFor() {