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 org.opendaylight.controller.cluster.raft.behaviors.Candidate;
23 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
24 import org.opendaylight.controller.cluster.raft.behaviors.Leader;
25 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
26 import org.opendaylight.controller.cluster.raft.client.messages.AddRaftPeer;
27 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
28 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
29 import org.opendaylight.controller.cluster.raft.internal.messages.ApplySnapshot;
30 import org.opendaylight.controller.cluster.raft.client.messages.RemoveRaftPeer;
31 import org.opendaylight.controller.cluster.raft.internal.messages.ApplyState;
32 import org.opendaylight.controller.cluster.raft.internal.messages.Replicate;
34 import java.io.Serializable;
35 import java.util.ArrayList;
36 import java.util.List;
40 * RaftActor encapsulates a state machine that needs to be kept synchronized
41 * in a cluster. It implements the RAFT algorithm as described in the paper
42 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
43 * In Search of an Understandable Consensus Algorithm</a>
45 * RaftActor has 3 states and each state has a certain behavior associated
46 * with it. A Raft actor can behave as,
49 * <li> A Follower (or) </li>
50 * <li> A Candidate </li>
54 * A RaftActor MUST be a Leader in order to accept requests from clients to
55 * change the state of it's encapsulated state machine. Once a RaftActor becomes
56 * a Leader it is also responsible for ensuring that all followers ultimately
57 * have the same log and therefore the same state machine as itself.
60 * The current behavior of a RaftActor determines how election for leadership
61 * is initiated and how peer RaftActors react to request for votes.
64 * Each RaftActor also needs to know the current election term. It uses this
65 * information for a couple of things. One is to simply figure out who it
66 * voted for in the last election. Another is to figure out if the message
67 * it received to update it's state is stale.
70 * The RaftActor uses akka-persistence to store it's replicated log.
71 * Furthermore through it's behaviors a Raft Actor determines
74 * <li> when a log entry should be persisted </li>
75 * <li> when a log entry should be applied to the state machine (and) </li>
76 * <li> when a snapshot should be saved </li>
79 public abstract class RaftActor extends UntypedPersistentActor {
80 protected final LoggingAdapter LOG =
81 Logging.getLogger(getContext().system(), this);
84 * The current state determines the current behavior of a RaftActor
85 * A Raft Actor always starts off in the Follower State
87 private RaftActorBehavior currentBehavior;
90 * This context should NOT be passed directly to any other actor it is
91 * only to be consumed by the RaftActorBehaviors
93 private RaftActorContext context;
96 * The in-memory journal
98 private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
101 public RaftActor(String id, Map<String, String> peerAddresses) {
102 context = new RaftActorContextImpl(this.getSelf(),
104 id, new ElectionTermImpl(getSelf().path().toString()),
105 -1, -1, replicatedLog, peerAddresses, LOG);
108 @Override public void onReceiveRecover(Object message) {
109 if (message instanceof SnapshotOffer) {
110 SnapshotOffer offer = (SnapshotOffer) message;
111 Snapshot snapshot = (Snapshot) offer.snapshot();
113 // Create a replicated log with the snapshot information
114 // The replicated log can be used later on to retrieve this snapshot
115 // when we need to install it on a peer
116 replicatedLog = new ReplicatedLogImpl(snapshot);
118 // Apply the snapshot to the actors state
119 applySnapshot(snapshot.getState());
121 } else if (message instanceof ReplicatedLogEntry) {
122 replicatedLog.append((ReplicatedLogEntry) message);
123 } else if (message instanceof RecoveryCompleted) {
125 "Last index in log : " + replicatedLog.lastIndex());
126 currentBehavior = switchBehavior(RaftState.Follower);
130 @Override public void onReceiveCommand(Object message) {
131 if (message instanceof ApplyState){
132 ApplyState applyState = (ApplyState) message;
134 LOG.debug("Applying state for log index {} data {}",
135 applyState.getReplicatedLogEntry().getIndex(),
136 applyState.getReplicatedLogEntry().getData());
138 applyState(applyState.getClientActor(), applyState.getIdentifier(),
139 applyState.getReplicatedLogEntry().getData());
140 } else if(message instanceof ApplySnapshot ) {
141 applySnapshot(((ApplySnapshot) message).getSnapshot());
142 } else if (message instanceof FindLeader) {
145 context.getPeerAddress(currentBehavior.getLeaderId())),
148 } else if (message instanceof SaveSnapshotSuccess) {
149 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
151 // TODO: Not sure if we want to be this aggressive with trimming stuff
152 trimPersistentData(success.metadata().sequenceNr());
154 } else if (message instanceof SaveSnapshotFailure) {
155 // TODO: Handle failure in saving the snapshot
156 } else if (message instanceof FindLeader){
157 getSender().tell(new FindLeaderReply(
158 context.getPeerAddress(currentBehavior.getLeaderId())),
161 } else if (message instanceof AddRaftPeer){
162 AddRaftPeer arp = (AddRaftPeer)message;
163 context.addToPeers(arp.getName(), arp.getAddress());
165 } else if (message instanceof RemoveRaftPeer){
166 RemoveRaftPeer rrp = (RemoveRaftPeer)message;
167 context.removePeer(rrp.getName());
170 currentBehavior.handleMessage(getSender(), message);
171 currentBehavior = switchBehavior(state);
178 * When a derived RaftActor needs to persist something it must call
185 protected void persistData(ActorRef clientActor, String identifier,
188 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
189 context.getReplicatedLog().lastIndex() + 1,
190 context.getTermInformation().getCurrentTerm(), data);
192 LOG.debug("Persist data {}", replicatedLogEntry);
195 .appendAndPersist(clientActor, identifier, replicatedLogEntry);
198 protected String getId() {
199 return context.getId();
203 * Derived actors can call the isLeader method to check if the current
204 * RaftActor is the Leader or not
206 * @return true it this RaftActor is a Leader false otherwise
208 protected boolean isLeader() {
209 return context.getId().equals(currentBehavior.getLeaderId());
213 * Derived actor can call getLeader if they need a reference to the Leader.
214 * This would be useful for example in forwarding a request to an actor
215 * which is the leader
217 * @return A reference to the leader if known, null otherwise
219 protected ActorSelection getLeader(){
220 String leaderId = currentBehavior.getLeaderId();
221 if (leaderId == null) {
224 String peerAddress = context.getPeerAddress(leaderId);
225 LOG.debug("getLeader leaderId = " + leaderId + " peerAddress = "
227 return context.actorSelection(peerAddress);
230 protected RaftState getRaftState() {
231 return currentBehavior.state();
237 * The applyState method will be called by the RaftActor when some data
238 * needs to be applied to the actor's state
240 * @param clientActor A reference to the client who sent this message. This
241 * is the same reference that was passed to persistData
242 * by the derived actor. clientActor may be null when
243 * the RaftActor is behaving as a follower or during
245 * @param identifier The identifier of the persisted data. This is also
246 * the same identifier that was passed to persistData by
247 * the derived actor. identifier may be null when
248 * the RaftActor is behaving as a follower or during
250 * @param data A piece of data that was persisted by the persistData call.
251 * This should NEVER be null.
253 protected abstract void applyState(ActorRef clientActor, String identifier,
257 * This method will be called by the RaftActor when a snapshot needs to be
258 * created. The derived actor should respond with its current state.
260 * During recovery the state that is returned by the derived actor will
261 * be passed back to it by calling the applySnapshot method
263 * @return The current state of the actor
265 protected abstract Object createSnapshot();
268 * This method will be called by the RaftActor during recovery to
269 * reconstruct the state of the actor.
271 * This method may also be called at any other point during normal
272 * operations when the derived actor is out of sync with it's peers
273 * and the only way to bring it in sync is by applying a snapshot
275 * @param snapshot A snapshot of the state of the actor
277 protected abstract void applySnapshot(Object snapshot);
279 private RaftActorBehavior switchBehavior(RaftState state) {
280 if (currentBehavior != null) {
281 if (currentBehavior.state() == state) {
282 return currentBehavior;
284 LOG.info("Switching from state " + currentBehavior.state() + " to "
288 currentBehavior.close();
289 } catch (Exception e) {
291 "Failed to close behavior : " + currentBehavior.state());
295 LOG.info("Switching behavior to " + state);
297 RaftActorBehavior behavior = null;
298 if (state == RaftState.Candidate) {
299 behavior = new Candidate(context);
300 } else if (state == RaftState.Follower) {
301 behavior = new Follower(context);
303 behavior = new Leader(context);
308 private void trimPersistentData(long sequenceNumber) {
310 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
311 // For now guessing that it is ANDed.
312 deleteSnapshots(new SnapshotSelectionCriteria(
313 sequenceNumber - 100000, 43200000));
316 deleteMessages(sequenceNumber);
319 private class ReplicatedLogImpl implements ReplicatedLog {
320 private final List<ReplicatedLogEntry> journal;
321 private final Object snapshot;
322 private long snapshotIndex = -1;
323 private long snapshotTerm = -1;
325 public ReplicatedLogImpl(Snapshot snapshot) {
326 this.snapshot = snapshot.getState();
327 this.snapshotIndex = snapshot.getLastAppliedIndex();
328 this.snapshotTerm = snapshot.getLastAppliedTerm();
330 this.journal = new ArrayList<>(snapshot.getUnAppliedEntries());
333 public ReplicatedLogImpl() {
334 this.snapshot = null;
335 this.journal = new ArrayList<>();
338 @Override public ReplicatedLogEntry get(long index) {
339 int adjustedIndex = adjustedIndex(index);
341 if (adjustedIndex < 0 || adjustedIndex >= journal.size()) {
345 return journal.get(adjustedIndex);
348 @Override public ReplicatedLogEntry last() {
349 if (journal.size() == 0) {
352 return get(journal.size() - 1);
355 @Override public long lastIndex() {
356 if (journal.size() == 0) {
360 return last().getIndex();
363 @Override public long lastTerm() {
364 if (journal.size() == 0) {
368 return last().getTerm();
372 @Override public void removeFrom(long index) {
373 int adjustedIndex = adjustedIndex(index);
375 if (adjustedIndex < 0 || adjustedIndex >= journal.size()) {
378 for (int i = adjustedIndex;
379 i < journal.size(); i++) {
385 @Override public void append(
386 final ReplicatedLogEntry replicatedLogEntry) {
387 journal.add(replicatedLogEntry);
390 @Override public List<ReplicatedLogEntry> getFrom(long index) {
391 int adjustedIndex = adjustedIndex(index);
393 List<ReplicatedLogEntry> entries = new ArrayList<>(100);
394 if (adjustedIndex < 0 || adjustedIndex >= journal.size()) {
397 for (int i = adjustedIndex;
398 i < journal.size(); i++) {
399 entries.add(journal.get(i));
404 @Override public void appendAndPersist(
405 final ReplicatedLogEntry replicatedLogEntry) {
406 appendAndPersist(null, null, replicatedLogEntry);
409 public void appendAndPersist(final ActorRef clientActor,
410 final String identifier,
411 final ReplicatedLogEntry replicatedLogEntry) {
412 context.getLogger().debug(
413 "Append log entry and persist " + replicatedLogEntry);
414 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
415 journal.add(replicatedLogEntry);
417 // When persisting events with persist it is guaranteed that the
418 // persistent actor will not receive further commands between the
419 // persist call and the execution(s) of the associated event
420 // handler. This also holds for multiple persist calls in context
421 // of a single command.
422 persist(replicatedLogEntry,
423 new Procedure<ReplicatedLogEntry>() {
424 public void apply(ReplicatedLogEntry evt) throws Exception {
425 // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned.
426 if (size() > 100000) {
427 ReplicatedLogEntry lastAppliedEntry =
428 get(context.getLastApplied());
429 long lastAppliedIndex = -1;
430 long lastAppliedTerm = -1;
431 if (lastAppliedEntry != null) {
432 lastAppliedIndex = lastAppliedEntry.getIndex();
433 lastAppliedTerm = lastAppliedEntry.getTerm();
436 saveSnapshot(Snapshot.create(createSnapshot(),
437 getFrom(context.getLastApplied() + 1),
438 lastIndex(), lastTerm(), lastAppliedIndex,
441 // Send message for replication
442 if (clientActor != null) {
443 currentBehavior.handleMessage(getSelf(),
444 new Replicate(clientActor, identifier,
453 @Override public long size() {
454 return journal.size() + snapshotIndex + 1;
457 @Override public boolean isPresent(long index) {
458 int adjustedIndex = adjustedIndex(index);
460 if (adjustedIndex < 0 || adjustedIndex >= journal.size()) {
466 @Override public boolean isInSnapshot(long index) {
467 return index <= snapshotIndex;
470 @Override public Object getSnapshot() {
474 @Override public long getSnapshotIndex() {
475 return snapshotIndex;
478 @Override public long getSnapshotTerm() {
482 private int adjustedIndex(long index) {
483 if(snapshotIndex < 0){
486 return (int) (index - snapshotIndex);
491 private static class ReplicatedLogImplEntry implements ReplicatedLogEntry,
494 private final long index;
495 private final long term;
496 private final Object payload;
498 public ReplicatedLogImplEntry(long index, long term, Object payload) {
502 this.payload = payload;
505 @Override public Object getData() {
509 @Override public long getTerm() {
513 @Override public long getIndex() {
517 @Override public String toString() {
526 private static class Snapshot implements Serializable {
527 private final Object state;
528 private final List<ReplicatedLogEntry> unAppliedEntries;
529 private final long lastIndex;
530 private final long lastTerm;
531 private final long lastAppliedIndex;
532 private final long lastAppliedTerm;
534 private Snapshot(Object state,
535 List<ReplicatedLogEntry> unAppliedEntries, long lastIndex,
536 long lastTerm, long lastAppliedIndex, long lastAppliedTerm) {
538 this.unAppliedEntries = unAppliedEntries;
539 this.lastIndex = lastIndex;
540 this.lastTerm = lastTerm;
541 this.lastAppliedIndex = lastAppliedIndex;
542 this.lastAppliedTerm = lastAppliedTerm;
546 public static Snapshot create(Object state,
547 List<ReplicatedLogEntry> entries, long lastIndex, long lastTerm,
548 long lastAppliedIndex, long lastAppliedTerm) {
549 return new Snapshot(state, entries, lastIndex, lastTerm,
550 lastAppliedIndex, lastAppliedTerm);
553 public Object getState() {
557 public List<ReplicatedLogEntry> getUnAppliedEntries() {
558 return unAppliedEntries;
561 public long getLastTerm() {
565 public long getLastAppliedIndex() {
566 return lastAppliedIndex;
569 public long getLastAppliedTerm() {
570 return lastAppliedTerm;