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);
154 @Override public void onReceiveCommand(Object message) {
155 if (message instanceof ApplyState){
156 ApplyState applyState = (ApplyState) message;
158 LOG.debug("Applying state for log index {} data {}",
159 applyState.getReplicatedLogEntry().getIndex(),
160 applyState.getReplicatedLogEntry().getData());
162 applyState(applyState.getClientActor(), applyState.getIdentifier(),
163 applyState.getReplicatedLogEntry().getData());
165 } else if(message instanceof ApplySnapshot ) {
166 applySnapshot(((ApplySnapshot) message).getSnapshot());
168 } else if (message instanceof FindLeader) {
171 context.getPeerAddress(currentBehavior.getLeaderId())),
175 } else if (message instanceof SaveSnapshotSuccess) {
176 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
178 // TODO: Not sure if we want to be this aggressive with trimming stuff
179 trimPersistentData(success.metadata().sequenceNr());
181 } else if (message instanceof SaveSnapshotFailure) {
183 // TODO: Handle failure in saving the snapshot
185 } else if (message instanceof FindLeader){
187 getSender().tell(new FindLeaderReply(
188 context.getPeerAddress(currentBehavior.getLeaderId())),
191 } else if (message instanceof AddRaftPeer){
193 // FIXME : Do not add raft peers like this.
194 // When adding a new Peer we have to ensure that the a majority of
195 // the peers know about the new Peer. Doing it this way may cause
196 // a situation where multiple Leaders may emerge
197 AddRaftPeer arp = (AddRaftPeer)message;
198 context.addToPeers(arp.getName(), arp.getAddress());
200 } else if (message instanceof RemoveRaftPeer){
202 RemoveRaftPeer rrp = (RemoveRaftPeer)message;
203 context.removePeer(rrp.getName());
208 currentBehavior.handleMessage(getSender(), message);
209 currentBehavior = switchBehavior(state);
213 public java.util.Set<String> getPeers() {
214 return context.getPeerAddresses().keySet();
217 protected String getReplicatedLogState() {
218 return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex()
219 + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm()
220 + ", im-mem journal size=" + context.getReplicatedLog().size();
225 * When a derived RaftActor needs to persist something it must call
232 protected void persistData(ActorRef clientActor, String identifier,
235 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
236 context.getReplicatedLog().lastIndex() + 1,
237 context.getTermInformation().getCurrentTerm(), data);
239 LOG.debug("Persist data {}", replicatedLogEntry);
242 .appendAndPersist(clientActor, identifier, replicatedLogEntry);
245 protected String getId() {
246 return context.getId();
250 * Derived actors can call the isLeader method to check if the current
251 * RaftActor is the Leader or not
253 * @return true it this RaftActor is a Leader false otherwise
255 protected boolean isLeader() {
256 return context.getId().equals(currentBehavior.getLeaderId());
260 * Derived actor can call getLeader if they need a reference to the Leader.
261 * This would be useful for example in forwarding a request to an actor
262 * which is the leader
264 * @return A reference to the leader if known, null otherwise
266 protected ActorSelection getLeader(){
267 String leaderId = currentBehavior.getLeaderId();
268 if (leaderId == null) {
271 String peerAddress = context.getPeerAddress(leaderId);
272 LOG.debug("getLeader leaderId = " + leaderId + " peerAddress = "
274 return context.actorSelection(peerAddress);
277 protected RaftState getRaftState() {
278 return currentBehavior.state();
282 * setPeerAddress sets the address of a known peer at a later time.
284 * This is to account for situations where a we know that a peer
285 * exists but we do not know an address up-front. This may also be used in
286 * situations where a known peer starts off in a different location and we
287 * need to change it's address
289 * Note that if the peerId does not match the list of peers passed to
290 * this actor during construction an IllegalStateException will be thrown.
295 protected void setPeerAddress(String peerId, String peerAddress){
296 context.setPeerAddress(peerId, peerAddress);
302 * The applyState method will be called by the RaftActor when some data
303 * needs to be applied to the actor's state
305 * @param clientActor A reference to the client who sent this message. This
306 * is the same reference that was passed to persistData
307 * by the derived actor. clientActor may be null when
308 * the RaftActor is behaving as a follower or during
310 * @param identifier The identifier of the persisted data. This is also
311 * the same identifier that was passed to persistData by
312 * the derived actor. identifier may be null when
313 * the RaftActor is behaving as a follower or during
315 * @param data A piece of data that was persisted by the persistData call.
316 * This should NEVER be null.
318 protected abstract void applyState(ActorRef clientActor, String identifier,
322 * This method will be called by the RaftActor when a snapshot needs to be
323 * created. The derived actor should respond with its current state.
325 * During recovery the state that is returned by the derived actor will
326 * be passed back to it by calling the applySnapshot method
328 * @return The current state of the actor
330 protected abstract Object createSnapshot();
333 * This method will be called by the RaftActor during recovery to
334 * reconstruct the state of the actor.
336 * This method may also be called at any other point during normal
337 * operations when the derived actor is out of sync with it's peers
338 * and the only way to bring it in sync is by applying a snapshot
340 * @param snapshot A snapshot of the state of the actor
342 protected abstract void applySnapshot(Object snapshot);
344 private RaftActorBehavior switchBehavior(RaftState state) {
345 if (currentBehavior != null) {
346 if (currentBehavior.state() == state) {
347 return currentBehavior;
349 LOG.info("Switching from state " + currentBehavior.state() + " to "
353 currentBehavior.close();
354 } catch (Exception e) {
356 "Failed to close behavior : " + currentBehavior.state());
360 LOG.info("Switching behavior to " + state);
362 RaftActorBehavior behavior = null;
363 if (state == RaftState.Candidate) {
364 behavior = new Candidate(context);
365 } else if (state == RaftState.Follower) {
366 behavior = new Follower(context);
368 behavior = new Leader(context);
373 private void trimPersistentData(long sequenceNumber) {
374 // Trim akka snapshots
375 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
376 // For now guessing that it is ANDed.
377 deleteSnapshots(new SnapshotSelectionCriteria(
378 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
381 deleteMessages(sequenceNumber);
385 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
387 public ReplicatedLogImpl(Snapshot snapshot) {
388 super(snapshot.getState(),
389 snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
390 snapshot.getUnAppliedEntries());
393 public ReplicatedLogImpl() {
397 @Override public void removeFromAndPersist(long logEntryIndex) {
398 int adjustedIndex = adjustedIndex(logEntryIndex);
400 if (adjustedIndex < 0) {
404 // FIXME: Maybe this should be done after the command is saved
405 journal.subList(adjustedIndex , journal.size()).clear();
407 persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
409 @Override public void apply(DeleteEntries param)
411 //FIXME : Doing nothing for now
416 @Override public void appendAndPersist(
417 final ReplicatedLogEntry replicatedLogEntry) {
418 appendAndPersist(null, null, replicatedLogEntry);
421 public void appendAndPersist(final ActorRef clientActor,
422 final String identifier,
423 final ReplicatedLogEntry replicatedLogEntry) {
424 context.getLogger().debug(
425 "Append log entry and persist {} ", replicatedLogEntry);
426 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
427 journal.add(replicatedLogEntry);
429 // When persisting events with persist it is guaranteed that the
430 // persistent actor will not receive further commands between the
431 // persist call and the execution(s) of the associated event
432 // handler. This also holds for multiple persist calls in context
433 // of a single command.
434 persist(replicatedLogEntry,
435 new Procedure<ReplicatedLogEntry>() {
436 public void apply(ReplicatedLogEntry evt) throws Exception {
437 // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned.
438 if (journal.size() > context.getConfigParams().getSnapshotBatchCount()) {
439 LOG.info("Initiating Snapshot Capture..");
440 long lastAppliedIndex = -1;
441 long lastAppliedTerm = -1;
443 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
444 if (lastAppliedEntry != null) {
445 lastAppliedIndex = lastAppliedEntry.getIndex();
446 lastAppliedTerm = lastAppliedEntry.getTerm();
449 LOG.debug("Snapshot Capture logSize: {}", journal.size());
450 LOG.debug("Snapshot Capture lastApplied:{} ", context.getLastApplied());
451 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
452 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
454 // create a snapshot object from the state provided and save it
455 // when snapshot is saved async, SaveSnapshotSuccess is raised.
456 Snapshot sn = Snapshot.create(createSnapshot(),
457 getFrom(context.getLastApplied() + 1),
458 lastIndex(), lastTerm(), lastAppliedIndex,
462 LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
464 //be greedy and remove entries from in-mem journal which are in the snapshot
465 // and update snapshotIndex and snapshotTerm without waiting for the success,
466 // TODO: damage-recovery to be done on failure
467 journal.subList(0, (int) (lastAppliedIndex - snapshotIndex)).clear();
468 snapshotIndex = lastAppliedIndex;
469 snapshotTerm = lastAppliedTerm;
471 LOG.info("Removed in-memory snapshotted entries, " +
472 "adjusted snaphsotIndex:{}" +
473 "and term:{}", snapshotIndex, lastAppliedTerm);
475 // Send message for replication
476 if (clientActor != null) {
477 currentBehavior.handleMessage(getSelf(),
478 new Replicate(clientActor, identifier,
489 private static class DeleteEntries implements Serializable {
490 private final int fromIndex;
493 public DeleteEntries(int fromIndex) {
494 this.fromIndex = fromIndex;
497 public int getFromIndex() {
503 private static class Snapshot implements Serializable {
504 private final Object state;
505 private final List<ReplicatedLogEntry> unAppliedEntries;
506 private final long lastIndex;
507 private final long lastTerm;
508 private final long lastAppliedIndex;
509 private final long lastAppliedTerm;
511 private Snapshot(Object state,
512 List<ReplicatedLogEntry> unAppliedEntries, long lastIndex,
513 long lastTerm, long lastAppliedIndex, long lastAppliedTerm) {
515 this.unAppliedEntries = unAppliedEntries;
516 this.lastIndex = lastIndex;
517 this.lastTerm = lastTerm;
518 this.lastAppliedIndex = lastAppliedIndex;
519 this.lastAppliedTerm = lastAppliedTerm;
523 public static Snapshot create(Object state,
524 List<ReplicatedLogEntry> entries, long lastIndex, long lastTerm,
525 long lastAppliedIndex, long lastAppliedTerm) {
526 return new Snapshot(state, entries, lastIndex, lastTerm,
527 lastAppliedIndex, lastAppliedTerm);
530 public Object getState() {
534 public List<ReplicatedLogEntry> getUnAppliedEntries() {
535 return unAppliedEntries;
538 public long getLastTerm() {
542 public long getLastAppliedIndex() {
543 return lastAppliedIndex;
546 public long getLastAppliedTerm() {
547 return lastAppliedTerm;
550 public String getLogMessage() {
551 StringBuilder sb = new StringBuilder();
552 return sb.append("Snapshot={")
553 .append("lastTerm:" + this.getLastTerm() + ", ")
554 .append("LastAppliedIndex:" + this.getLastAppliedIndex() + ", ")
555 .append("LastAppliedTerm:" + this.getLastAppliedTerm() + ", ")
556 .append("UnAppliedEntries size:" + this.getUnAppliedEntries().size() + "}")
562 private class ElectionTermImpl implements ElectionTerm {
564 * Identifier of the actor whose election term information this is
566 private long currentTerm = 0;
567 private String votedFor = null;
569 public long getCurrentTerm() {
573 public String getVotedFor() {
577 @Override public void update(long currentTerm, String votedFor) {
578 LOG.info("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
580 this.currentTerm = currentTerm;
581 this.votedFor = votedFor;
585 public void updateAndPersist(long currentTerm, String votedFor){
586 update(currentTerm, votedFor);
587 // FIXME : Maybe first persist then update the state
588 persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
590 @Override public void apply(UpdateElectionTerm param)
598 private static class UpdateElectionTerm implements Serializable {
599 private final long currentTerm;
600 private final String votedFor;
602 public UpdateElectionTerm(long currentTerm, String votedFor) {
603 this.currentTerm = currentTerm;
604 this.votedFor = votedFor;
607 public long getCurrentTerm() {
611 public String getVotedFor() {