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();
298 protected ReplicatedLogEntry getLastLogEntry() {
299 return replicatedLog.last();
302 protected Long getCurrentTerm(){
303 return context.getTermInformation().getCurrentTerm();
306 protected Long getCommitIndex(){
307 return context.getCommitIndex();
310 protected Long getLastApplied(){
311 return context.getLastApplied();
315 * setPeerAddress sets the address of a known peer at a later time.
317 * This is to account for situations where a we know that a peer
318 * exists but we do not know an address up-front. This may also be used in
319 * situations where a known peer starts off in a different location and we
320 * need to change it's address
322 * Note that if the peerId does not match the list of peers passed to
323 * this actor during construction an IllegalStateException will be thrown.
328 protected void setPeerAddress(String peerId, String peerAddress){
329 context.setPeerAddress(peerId, peerAddress);
335 * The applyState method will be called by the RaftActor when some data
336 * needs to be applied to the actor's state
338 * @param clientActor A reference to the client who sent this message. This
339 * is the same reference that was passed to persistData
340 * by the derived actor. clientActor may be null when
341 * the RaftActor is behaving as a follower or during
343 * @param identifier The identifier of the persisted data. This is also
344 * the same identifier that was passed to persistData by
345 * the derived actor. identifier may be null when
346 * the RaftActor is behaving as a follower or during
348 * @param data A piece of data that was persisted by the persistData call.
349 * This should NEVER be null.
351 protected abstract void applyState(ActorRef clientActor, String identifier,
355 * This method will be called by the RaftActor when a snapshot needs to be
356 * created. The derived actor should respond with its current state.
358 * During recovery the state that is returned by the derived actor will
359 * be passed back to it by calling the applySnapshot method
361 * @return The current state of the actor
363 protected abstract Object createSnapshot();
366 * This method will be called by the RaftActor during recovery to
367 * reconstruct the state of the actor.
369 * This method may also be called at any other point during normal
370 * operations when the derived actor is out of sync with it's peers
371 * and the only way to bring it in sync is by applying a snapshot
373 * @param snapshot A snapshot of the state of the actor
375 protected abstract void applySnapshot(Object snapshot);
378 * This method will be called by the RaftActor when the state of the
379 * RaftActor changes. The derived actor can then use methods like
380 * isLeader or getLeader to do something useful
382 protected abstract void onStateChanged();
384 private RaftActorBehavior switchBehavior(RaftState state) {
385 if (currentBehavior != null) {
386 if (currentBehavior.state() == state) {
387 return currentBehavior;
389 LOG.info("Switching from state " + currentBehavior.state() + " to "
393 currentBehavior.close();
394 } catch (Exception e) {
396 "Failed to close behavior : " + currentBehavior.state());
400 LOG.info("Switching behavior to " + state);
402 RaftActorBehavior behavior = null;
403 if (state == RaftState.Candidate) {
404 behavior = new Candidate(context);
405 } else if (state == RaftState.Follower) {
406 behavior = new Follower(context);
408 behavior = new Leader(context);
416 private void trimPersistentData(long sequenceNumber) {
417 // Trim akka snapshots
418 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
419 // For now guessing that it is ANDed.
420 deleteSnapshots(new SnapshotSelectionCriteria(
421 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
424 deleteMessages(sequenceNumber);
428 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
430 public ReplicatedLogImpl(Snapshot snapshot) {
431 super(snapshot.getState(),
432 snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
433 snapshot.getUnAppliedEntries());
436 public ReplicatedLogImpl() {
440 @Override public void removeFromAndPersist(long logEntryIndex) {
441 int adjustedIndex = adjustedIndex(logEntryIndex);
443 if (adjustedIndex < 0) {
447 // FIXME: Maybe this should be done after the command is saved
448 journal.subList(adjustedIndex , journal.size()).clear();
450 persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
452 @Override public void apply(DeleteEntries param)
454 //FIXME : Doing nothing for now
459 @Override public void appendAndPersist(
460 final ReplicatedLogEntry replicatedLogEntry) {
461 appendAndPersist(null, null, replicatedLogEntry);
464 public void appendAndPersist(final ActorRef clientActor,
465 final String identifier,
466 final ReplicatedLogEntry replicatedLogEntry) {
467 context.getLogger().debug(
468 "Append log entry and persist {} ", replicatedLogEntry);
469 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
470 journal.add(replicatedLogEntry);
472 // When persisting events with persist it is guaranteed that the
473 // persistent actor will not receive further commands between the
474 // persist call and the execution(s) of the associated event
475 // handler. This also holds for multiple persist calls in context
476 // of a single command.
477 persist(replicatedLogEntry,
478 new Procedure<ReplicatedLogEntry>() {
479 public void apply(ReplicatedLogEntry evt) throws Exception {
480 // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned.
481 if (journal.size() > context.getConfigParams().getSnapshotBatchCount()) {
482 LOG.info("Initiating Snapshot Capture..");
483 long lastAppliedIndex = -1;
484 long lastAppliedTerm = -1;
486 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
487 if (lastAppliedEntry != null) {
488 lastAppliedIndex = lastAppliedEntry.getIndex();
489 lastAppliedTerm = lastAppliedEntry.getTerm();
492 LOG.debug("Snapshot Capture logSize: {}", journal.size());
493 LOG.debug("Snapshot Capture lastApplied:{} ", context.getLastApplied());
494 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
495 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
497 // create a snapshot object from the state provided and save it
498 // when snapshot is saved async, SaveSnapshotSuccess is raised.
499 Snapshot sn = Snapshot.create(createSnapshot(),
500 getFrom(context.getLastApplied() + 1),
501 lastIndex(), lastTerm(), lastAppliedIndex,
505 LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
507 //be greedy and remove entries from in-mem journal which are in the snapshot
508 // and update snapshotIndex and snapshotTerm without waiting for the success,
509 // TODO: damage-recovery to be done on failure
510 journal.subList(0, (int) (lastAppliedIndex - snapshotIndex)).clear();
511 snapshotIndex = lastAppliedIndex;
512 snapshotTerm = lastAppliedTerm;
514 LOG.info("Removed in-memory snapshotted entries, " +
515 "adjusted snaphsotIndex:{}" +
516 "and term:{}", snapshotIndex, lastAppliedTerm);
518 // Send message for replication
519 if (clientActor != null) {
520 currentBehavior.handleMessage(getSelf(),
521 new Replicate(clientActor, identifier,
532 private static class DeleteEntries implements Serializable {
533 private final int fromIndex;
536 public DeleteEntries(int fromIndex) {
537 this.fromIndex = fromIndex;
540 public int getFromIndex() {
546 private static class Snapshot implements Serializable {
547 private final Object state;
548 private final List<ReplicatedLogEntry> unAppliedEntries;
549 private final long lastIndex;
550 private final long lastTerm;
551 private final long lastAppliedIndex;
552 private final long lastAppliedTerm;
554 private Snapshot(Object state,
555 List<ReplicatedLogEntry> unAppliedEntries, long lastIndex,
556 long lastTerm, long lastAppliedIndex, long lastAppliedTerm) {
558 this.unAppliedEntries = unAppliedEntries;
559 this.lastIndex = lastIndex;
560 this.lastTerm = lastTerm;
561 this.lastAppliedIndex = lastAppliedIndex;
562 this.lastAppliedTerm = lastAppliedTerm;
566 public static Snapshot create(Object state,
567 List<ReplicatedLogEntry> entries, long lastIndex, long lastTerm,
568 long lastAppliedIndex, long lastAppliedTerm) {
569 return new Snapshot(state, entries, lastIndex, lastTerm,
570 lastAppliedIndex, lastAppliedTerm);
573 public Object getState() {
577 public List<ReplicatedLogEntry> getUnAppliedEntries() {
578 return unAppliedEntries;
581 public long getLastTerm() {
585 public long getLastAppliedIndex() {
586 return lastAppliedIndex;
589 public long getLastAppliedTerm() {
590 return lastAppliedTerm;
593 public String getLogMessage() {
594 StringBuilder sb = new StringBuilder();
595 return sb.append("Snapshot={")
596 .append("lastTerm:" + this.getLastTerm() + ", ")
597 .append("LastAppliedIndex:" + this.getLastAppliedIndex() + ", ")
598 .append("LastAppliedTerm:" + this.getLastAppliedTerm() + ", ")
599 .append("UnAppliedEntries size:" + this.getUnAppliedEntries().size() + "}")
605 private class ElectionTermImpl implements ElectionTerm {
607 * Identifier of the actor whose election term information this is
609 private long currentTerm = 0;
610 private String votedFor = null;
612 public long getCurrentTerm() {
616 public String getVotedFor() {
620 @Override public void update(long currentTerm, String votedFor) {
621 LOG.debug("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
623 this.currentTerm = currentTerm;
624 this.votedFor = votedFor;
628 public void updateAndPersist(long currentTerm, String votedFor){
629 update(currentTerm, votedFor);
630 // FIXME : Maybe first persist then update the state
631 persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
633 @Override public void apply(UpdateElectionTerm param)
641 private static class UpdateElectionTerm implements Serializable {
642 private final long currentTerm;
643 private final String votedFor;
645 public UpdateElectionTerm(long currentTerm, String votedFor) {
646 this.currentTerm = currentTerm;
647 this.votedFor = votedFor;
650 public long getCurrentTerm() {
654 public String getVotedFor() {