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.base.messages.ApplySnapshot;
23 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
24 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
25 import com.google.common.base.Optional;
26 import org.opendaylight.controller.cluster.raft.behaviors.Candidate;
27 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
28 import org.opendaylight.controller.cluster.raft.behaviors.Leader;
29 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
30 import org.opendaylight.controller.cluster.raft.client.messages.AddRaftPeer;
31 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
32 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
33 import org.opendaylight.controller.cluster.raft.client.messages.RemoveRaftPeer;
34 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
36 import java.io.Serializable;
37 import java.util.List;
41 * RaftActor encapsulates a state machine that needs to be kept synchronized
42 * in a cluster. It implements the RAFT algorithm as described in the paper
43 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
44 * In Search of an Understandable Consensus Algorithm</a>
46 * RaftActor has 3 states and each state has a certain behavior associated
47 * with it. A Raft actor can behave as,
50 * <li> A Follower (or) </li>
51 * <li> A Candidate </li>
55 * A RaftActor MUST be a Leader in order to accept requests from clients to
56 * change the state of it's encapsulated state machine. Once a RaftActor becomes
57 * a Leader it is also responsible for ensuring that all followers ultimately
58 * have the same log and therefore the same state machine as itself.
61 * The current behavior of a RaftActor determines how election for leadership
62 * is initiated and how peer RaftActors react to request for votes.
65 * Each RaftActor also needs to know the current election term. It uses this
66 * information for a couple of things. One is to simply figure out who it
67 * voted for in the last election. Another is to figure out if the message
68 * it received to update it's state is stale.
71 * The RaftActor uses akka-persistence to store it's replicated log.
72 * Furthermore through it's behaviors a Raft Actor determines
75 * <li> when a log entry should be persisted </li>
76 * <li> when a log entry should be applied to the state machine (and) </li>
77 * <li> when a snapshot should be saved </li>
80 public abstract class RaftActor extends UntypedPersistentActor {
81 protected final LoggingAdapter LOG =
82 Logging.getLogger(getContext().system(), this);
85 * The current state determines the current behavior of a RaftActor
86 * A Raft Actor always starts off in the Follower State
88 private RaftActorBehavior currentBehavior;
91 * This context should NOT be passed directly to any other actor it is
92 * only to be consumed by the RaftActorBehaviors
94 private RaftActorContext context;
97 * The in-memory journal
99 private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
102 public RaftActor(String id, Map<String, String> peerAddresses) {
103 this(id, peerAddresses, Optional.<ConfigParams>absent());
106 public RaftActor(String id, Map<String, String> peerAddresses,
107 Optional<ConfigParams> configParams) {
109 context = new RaftActorContextImpl(this.getSelf(),
110 this.getContext(), id, new ElectionTermImpl(),
111 -1, -1, replicatedLog, peerAddresses,
112 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
116 @Override public void onReceiveRecover(Object message) {
117 if (message instanceof SnapshotOffer) {
118 LOG.debug("SnapshotOffer called..");
119 SnapshotOffer offer = (SnapshotOffer) message;
120 Snapshot snapshot = (Snapshot) offer.snapshot();
122 // Create a replicated log with the snapshot information
123 // The replicated log can be used later on to retrieve this snapshot
124 // when we need to install it on a peer
125 replicatedLog = new ReplicatedLogImpl(snapshot);
127 context.setReplicatedLog(replicatedLog);
129 LOG.debug("Applied snapshot to replicatedLog. " +
130 "snapshotIndex={}, snapshotTerm={}, journal-size={}",
131 replicatedLog.snapshotIndex, replicatedLog.snapshotTerm,
132 replicatedLog.size());
134 // Apply the snapshot to the actors state
135 applySnapshot(snapshot.getState());
137 } else if (message instanceof ReplicatedLogEntry) {
138 replicatedLog.append((ReplicatedLogEntry) message);
139 } else if (message instanceof DeleteEntries) {
140 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
141 } else if (message instanceof UpdateElectionTerm) {
142 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor());
143 } else if (message instanceof RecoveryCompleted) {
145 "RecoveryCompleted - Switching actor to Follower - " +
146 "Last index in log:{}, snapshotIndex={}, snapshotTerm={}, " +
148 replicatedLog.lastIndex(), replicatedLog.snapshotIndex,
149 replicatedLog.snapshotTerm, replicatedLog.size());
150 currentBehavior = switchBehavior(RaftState.Follower);
155 @Override public void onReceiveCommand(Object message) {
156 if (message instanceof ApplyState){
157 ApplyState applyState = (ApplyState) message;
159 LOG.debug("Applying state for log index {} data {}",
160 applyState.getReplicatedLogEntry().getIndex(),
161 applyState.getReplicatedLogEntry().getData());
163 applyState(applyState.getClientActor(), applyState.getIdentifier(),
164 applyState.getReplicatedLogEntry().getData());
166 } else if(message instanceof ApplySnapshot ) {
167 applySnapshot(((ApplySnapshot) message).getSnapshot());
169 } else if (message instanceof FindLeader) {
172 context.getPeerAddress(currentBehavior.getLeaderId())),
176 } else if (message instanceof SaveSnapshotSuccess) {
177 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
179 // TODO: Not sure if we want to be this aggressive with trimming stuff
180 trimPersistentData(success.metadata().sequenceNr());
182 } else if (message instanceof SaveSnapshotFailure) {
184 // TODO: Handle failure in saving the snapshot
186 } else if (message instanceof FindLeader){
188 getSender().tell(new FindLeaderReply(
189 context.getPeerAddress(currentBehavior.getLeaderId())),
192 } else if (message instanceof AddRaftPeer){
194 // FIXME : Do not add raft peers like this.
195 // When adding a new Peer we have to ensure that the a majority of
196 // the peers know about the new Peer. Doing it this way may cause
197 // a situation where multiple Leaders may emerge
198 AddRaftPeer arp = (AddRaftPeer)message;
199 context.addToPeers(arp.getName(), arp.getAddress());
201 } else if (message instanceof RemoveRaftPeer){
203 RemoveRaftPeer rrp = (RemoveRaftPeer)message;
204 context.removePeer(rrp.getName());
209 currentBehavior.handleMessage(getSender(), message);
210 RaftActorBehavior oldBehavior = currentBehavior;
211 currentBehavior = switchBehavior(state);
212 if(oldBehavior != currentBehavior){
218 public java.util.Set<String> getPeers() {
219 return context.getPeerAddresses().keySet();
222 protected String getReplicatedLogState() {
223 return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex()
224 + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm()
225 + ", im-mem journal size=" + context.getReplicatedLog().size();
230 * When a derived RaftActor needs to persist something it must call
237 protected void persistData(ActorRef clientActor, String identifier,
240 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
241 context.getReplicatedLog().lastIndex() + 1,
242 context.getTermInformation().getCurrentTerm(), data);
244 LOG.debug("Persist data {}", replicatedLogEntry);
247 .appendAndPersist(clientActor, identifier, replicatedLogEntry);
250 protected String getId() {
251 return context.getId();
255 * Derived actors can call the isLeader method to check if the current
256 * RaftActor is the Leader or not
258 * @return true it this RaftActor is a Leader false otherwise
260 protected boolean isLeader() {
261 return context.getId().equals(currentBehavior.getLeaderId());
265 * Derived actor can call getLeader if they need a reference to the Leader.
266 * This would be useful for example in forwarding a request to an actor
267 * which is the leader
269 * @return A reference to the leader if known, null otherwise
271 protected ActorSelection getLeader(){
272 String leaderId = currentBehavior.getLeaderId();
273 if (leaderId == null) {
276 String peerAddress = context.getPeerAddress(leaderId);
277 LOG.debug("getLeader leaderId = " + leaderId + " peerAddress = "
280 if(peerAddress == null){
283 return context.actorSelection(peerAddress);
288 * @return the current leader's id
290 protected String getLeaderId(){
291 return currentBehavior.getLeaderId();
294 protected RaftState getRaftState() {
295 return currentBehavior.state();
299 * setPeerAddress sets the address of a known peer at a later time.
301 * This is to account for situations where a we know that a peer
302 * exists but we do not know an address up-front. This may also be used in
303 * situations where a known peer starts off in a different location and we
304 * need to change it's address
306 * Note that if the peerId does not match the list of peers passed to
307 * this actor during construction an IllegalStateException will be thrown.
312 protected void setPeerAddress(String peerId, String peerAddress){
313 context.setPeerAddress(peerId, peerAddress);
319 * The applyState method will be called by the RaftActor when some data
320 * needs to be applied to the actor's state
322 * @param clientActor A reference to the client who sent this message. This
323 * is the same reference that was passed to persistData
324 * by the derived actor. clientActor may be null when
325 * the RaftActor is behaving as a follower or during
327 * @param identifier The identifier of the persisted data. This is also
328 * the same identifier that was passed to persistData by
329 * the derived actor. identifier may be null when
330 * the RaftActor is behaving as a follower or during
332 * @param data A piece of data that was persisted by the persistData call.
333 * This should NEVER be null.
335 protected abstract void applyState(ActorRef clientActor, String identifier,
339 * This method will be called by the RaftActor when a snapshot needs to be
340 * created. The derived actor should respond with its current state.
342 * During recovery the state that is returned by the derived actor will
343 * be passed back to it by calling the applySnapshot method
345 * @return The current state of the actor
347 protected abstract Object createSnapshot();
350 * This method will be called by the RaftActor during recovery to
351 * reconstruct the state of the actor.
353 * This method may also be called at any other point during normal
354 * operations when the derived actor is out of sync with it's peers
355 * and the only way to bring it in sync is by applying a snapshot
357 * @param snapshot A snapshot of the state of the actor
359 protected abstract void applySnapshot(Object snapshot);
362 * This method will be called by the RaftActor when the state of the
363 * RaftActor changes. The derived actor can then use methods like
364 * isLeader or getLeader to do something useful
366 protected abstract void onStateChanged();
368 private RaftActorBehavior switchBehavior(RaftState state) {
369 if (currentBehavior != null) {
370 if (currentBehavior.state() == state) {
371 return currentBehavior;
373 LOG.info("Switching from state " + currentBehavior.state() + " to "
377 currentBehavior.close();
378 } catch (Exception e) {
380 "Failed to close behavior : " + currentBehavior.state());
384 LOG.info("Switching behavior to " + state);
386 RaftActorBehavior behavior = null;
387 if (state == RaftState.Candidate) {
388 behavior = new Candidate(context);
389 } else if (state == RaftState.Follower) {
390 behavior = new Follower(context);
392 behavior = new Leader(context);
400 private void trimPersistentData(long sequenceNumber) {
401 // Trim akka snapshots
402 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
403 // For now guessing that it is ANDed.
404 deleteSnapshots(new SnapshotSelectionCriteria(
405 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
408 deleteMessages(sequenceNumber);
412 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
414 public ReplicatedLogImpl(Snapshot snapshot) {
415 super(snapshot.getState(),
416 snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
417 snapshot.getUnAppliedEntries());
420 public ReplicatedLogImpl() {
424 @Override public void removeFromAndPersist(long logEntryIndex) {
425 int adjustedIndex = adjustedIndex(logEntryIndex);
427 if (adjustedIndex < 0) {
431 // FIXME: Maybe this should be done after the command is saved
432 journal.subList(adjustedIndex , journal.size()).clear();
434 persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
436 @Override public void apply(DeleteEntries param)
438 //FIXME : Doing nothing for now
443 @Override public void appendAndPersist(
444 final ReplicatedLogEntry replicatedLogEntry) {
445 appendAndPersist(null, null, replicatedLogEntry);
448 public void appendAndPersist(final ActorRef clientActor,
449 final String identifier,
450 final ReplicatedLogEntry replicatedLogEntry) {
451 context.getLogger().debug(
452 "Append log entry and persist {} ", replicatedLogEntry);
453 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
454 journal.add(replicatedLogEntry);
456 // When persisting events with persist it is guaranteed that the
457 // persistent actor will not receive further commands between the
458 // persist call and the execution(s) of the associated event
459 // handler. This also holds for multiple persist calls in context
460 // of a single command.
461 persist(replicatedLogEntry,
462 new Procedure<ReplicatedLogEntry>() {
463 public void apply(ReplicatedLogEntry evt) throws Exception {
464 // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned.
465 if (journal.size() > context.getConfigParams().getSnapshotBatchCount()) {
466 LOG.info("Initiating Snapshot Capture..");
467 long lastAppliedIndex = -1;
468 long lastAppliedTerm = -1;
470 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
471 if (lastAppliedEntry != null) {
472 lastAppliedIndex = lastAppliedEntry.getIndex();
473 lastAppliedTerm = lastAppliedEntry.getTerm();
476 LOG.debug("Snapshot Capture logSize: {}", journal.size());
477 LOG.debug("Snapshot Capture lastApplied:{} ", context.getLastApplied());
478 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
479 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
481 // create a snapshot object from the state provided and save it
482 // when snapshot is saved async, SaveSnapshotSuccess is raised.
483 Snapshot sn = Snapshot.create(createSnapshot(),
484 getFrom(context.getLastApplied() + 1),
485 lastIndex(), lastTerm(), lastAppliedIndex,
489 LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
491 //be greedy and remove entries from in-mem journal which are in the snapshot
492 // and update snapshotIndex and snapshotTerm without waiting for the success,
493 // TODO: damage-recovery to be done on failure
494 journal.subList(0, (int) (lastAppliedIndex - snapshotIndex)).clear();
495 snapshotIndex = lastAppliedIndex;
496 snapshotTerm = lastAppliedTerm;
498 LOG.info("Removed in-memory snapshotted entries, " +
499 "adjusted snaphsotIndex:{}" +
500 "and term:{}", snapshotIndex, lastAppliedTerm);
502 // Send message for replication
503 if (clientActor != null) {
504 currentBehavior.handleMessage(getSelf(),
505 new Replicate(clientActor, identifier,
516 private static class DeleteEntries implements Serializable {
517 private final int fromIndex;
520 public DeleteEntries(int fromIndex) {
521 this.fromIndex = fromIndex;
524 public int getFromIndex() {
530 private static class Snapshot implements Serializable {
531 private final Object state;
532 private final List<ReplicatedLogEntry> unAppliedEntries;
533 private final long lastIndex;
534 private final long lastTerm;
535 private final long lastAppliedIndex;
536 private final long lastAppliedTerm;
538 private Snapshot(Object state,
539 List<ReplicatedLogEntry> unAppliedEntries, long lastIndex,
540 long lastTerm, long lastAppliedIndex, long lastAppliedTerm) {
542 this.unAppliedEntries = unAppliedEntries;
543 this.lastIndex = lastIndex;
544 this.lastTerm = lastTerm;
545 this.lastAppliedIndex = lastAppliedIndex;
546 this.lastAppliedTerm = lastAppliedTerm;
550 public static Snapshot create(Object state,
551 List<ReplicatedLogEntry> entries, long lastIndex, long lastTerm,
552 long lastAppliedIndex, long lastAppliedTerm) {
553 return new Snapshot(state, entries, lastIndex, lastTerm,
554 lastAppliedIndex, lastAppliedTerm);
557 public Object getState() {
561 public List<ReplicatedLogEntry> getUnAppliedEntries() {
562 return unAppliedEntries;
565 public long getLastTerm() {
569 public long getLastAppliedIndex() {
570 return lastAppliedIndex;
573 public long getLastAppliedTerm() {
574 return lastAppliedTerm;
577 public String getLogMessage() {
578 StringBuilder sb = new StringBuilder();
579 return sb.append("Snapshot={")
580 .append("lastTerm:" + this.getLastTerm() + ", ")
581 .append("LastAppliedIndex:" + this.getLastAppliedIndex() + ", ")
582 .append("LastAppliedTerm:" + this.getLastAppliedTerm() + ", ")
583 .append("UnAppliedEntries size:" + this.getUnAppliedEntries().size() + "}")
589 private class ElectionTermImpl implements ElectionTerm {
591 * Identifier of the actor whose election term information this is
593 private long currentTerm = 0;
594 private String votedFor = null;
596 public long getCurrentTerm() {
600 public String getVotedFor() {
604 @Override public void update(long currentTerm, String votedFor) {
605 LOG.info("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
607 this.currentTerm = currentTerm;
608 this.votedFor = votedFor;
612 public void updateAndPersist(long currentTerm, String votedFor){
613 update(currentTerm, votedFor);
614 // FIXME : Maybe first persist then update the state
615 persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
617 @Override public void apply(UpdateElectionTerm param)
625 private static class UpdateElectionTerm implements Serializable {
626 private final long currentTerm;
627 private final String votedFor;
629 public UpdateElectionTerm(long currentTerm, String votedFor) {
630 this.currentTerm = currentTerm;
631 this.votedFor = votedFor;
634 public long getCurrentTerm() {
638 public String getVotedFor() {