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 org.opendaylight.controller.cluster.raft.behaviors.Candidate;
26 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
27 import org.opendaylight.controller.cluster.raft.behaviors.Leader;
28 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
29 import org.opendaylight.controller.cluster.raft.client.messages.AddRaftPeer;
30 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
31 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
32 import org.opendaylight.controller.cluster.raft.client.messages.RemoveRaftPeer;
33 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
35 import java.io.Serializable;
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);
83 private static final int SNAPSHOT_ENTRY_COUNT = 100000;
86 * The current state determines the current behavior of a RaftActor
87 * A Raft Actor always starts off in the Follower State
89 private RaftActorBehavior currentBehavior;
92 * This context should NOT be passed directly to any other actor it is
93 * only to be consumed by the RaftActorBehaviors
95 private RaftActorContext context;
98 * The in-memory journal
100 private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
103 public RaftActor(String id, Map<String, String> peerAddresses) {
104 context = new RaftActorContextImpl(this.getSelf(),
106 id, new ElectionTermImpl(),
107 -1, -1, replicatedLog, peerAddresses, LOG);
110 @Override public void onReceiveRecover(Object message) {
111 if (message instanceof SnapshotOffer) {
112 LOG.debug("SnapshotOffer called..");
113 SnapshotOffer offer = (SnapshotOffer) message;
114 Snapshot snapshot = (Snapshot) offer.snapshot();
116 // Create a replicated log with the snapshot information
117 // The replicated log can be used later on to retrieve this snapshot
118 // when we need to install it on a peer
119 replicatedLog = new ReplicatedLogImpl(snapshot);
121 context.setReplicatedLog(replicatedLog);
123 LOG.debug("Applied snapshot to replicatedLog. " +
124 "snapshotIndex={}, snapshotTerm={}, journal-size={}",
125 replicatedLog.snapshotIndex, replicatedLog.snapshotTerm,
126 replicatedLog.size());
128 // Apply the snapshot to the actors state
129 applySnapshot(snapshot.getState());
131 } else if (message instanceof ReplicatedLogEntry) {
132 replicatedLog.append((ReplicatedLogEntry) message);
133 } else if (message instanceof DeleteEntries) {
134 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
135 } else if (message instanceof UpdateElectionTerm) {
136 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor());
137 } else if (message instanceof RecoveryCompleted) {
139 "RecoveryCompleted - Switching actor to Follower - " +
140 "Last index in log:{}, snapshotIndex={}, snapshotTerm={}, " +
142 replicatedLog.lastIndex(), replicatedLog.snapshotIndex,
143 replicatedLog.snapshotTerm, replicatedLog.size());
144 currentBehavior = switchBehavior(RaftState.Follower);
148 @Override public void onReceiveCommand(Object message) {
149 if (message instanceof ApplyState){
150 ApplyState applyState = (ApplyState) message;
152 LOG.debug("Applying state for log index {} data {}",
153 applyState.getReplicatedLogEntry().getIndex(),
154 applyState.getReplicatedLogEntry().getData());
156 applyState(applyState.getClientActor(), applyState.getIdentifier(),
157 applyState.getReplicatedLogEntry().getData());
159 } else if(message instanceof ApplySnapshot ) {
160 applySnapshot(((ApplySnapshot) message).getSnapshot());
162 } else if (message instanceof FindLeader) {
165 context.getPeerAddress(currentBehavior.getLeaderId())),
169 } else if (message instanceof SaveSnapshotSuccess) {
170 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
172 // TODO: Not sure if we want to be this aggressive with trimming stuff
173 trimPersistentData(success.metadata().sequenceNr());
175 } else if (message instanceof SaveSnapshotFailure) {
177 // TODO: Handle failure in saving the snapshot
179 } else if (message instanceof FindLeader){
181 getSender().tell(new FindLeaderReply(
182 context.getPeerAddress(currentBehavior.getLeaderId())),
185 } else if (message instanceof AddRaftPeer){
187 // FIXME : Do not add raft peers like this.
188 // When adding a new Peer we have to ensure that the a majority of
189 // the peers know about the new Peer. Doing it this way may cause
190 // a situation where multiple Leaders may emerge
191 AddRaftPeer arp = (AddRaftPeer)message;
192 context.addToPeers(arp.getName(), arp.getAddress());
194 } else if (message instanceof RemoveRaftPeer){
196 RemoveRaftPeer rrp = (RemoveRaftPeer)message;
197 context.removePeer(rrp.getName());
202 currentBehavior.handleMessage(getSender(), message);
203 currentBehavior = switchBehavior(state);
207 public java.util.Set<String> getPeers() {
208 return context.getPeerAddresses().keySet();
211 protected String getReplicatedLogState() {
212 return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex()
213 + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm()
214 + ", im-mem journal size=" + context.getReplicatedLog().size();
219 * When a derived RaftActor needs to persist something it must call
226 protected void persistData(ActorRef clientActor, String identifier,
229 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
230 context.getReplicatedLog().lastIndex() + 1,
231 context.getTermInformation().getCurrentTerm(), data);
233 LOG.debug("Persist data {}", replicatedLogEntry);
236 .appendAndPersist(clientActor, identifier, replicatedLogEntry);
239 protected String getId() {
240 return context.getId();
244 * Derived actors can call the isLeader method to check if the current
245 * RaftActor is the Leader or not
247 * @return true it this RaftActor is a Leader false otherwise
249 protected boolean isLeader() {
250 return context.getId().equals(currentBehavior.getLeaderId());
254 * Derived actor can call getLeader if they need a reference to the Leader.
255 * This would be useful for example in forwarding a request to an actor
256 * which is the leader
258 * @return A reference to the leader if known, null otherwise
260 protected ActorSelection getLeader(){
261 String leaderId = currentBehavior.getLeaderId();
262 if (leaderId == null) {
265 String peerAddress = context.getPeerAddress(leaderId);
266 LOG.debug("getLeader leaderId = " + leaderId + " peerAddress = "
268 return context.actorSelection(peerAddress);
271 protected RaftState getRaftState() {
272 return currentBehavior.state();
276 * setPeerAddress sets the address of a known peer at a later time.
278 * This is to account for situations where a we know that a peer
279 * exists but we do not know an address up-front. This may also be used in
280 * situations where a known peer starts off in a different location and we
281 * need to change it's address
283 * Note that if the peerId does not match the list of peers passed to
284 * this actor during construction an IllegalStateException will be thrown.
289 protected void setPeerAddress(String peerId, String peerAddress){
290 context.setPeerAddress(peerId, peerAddress);
296 * The applyState method will be called by the RaftActor when some data
297 * needs to be applied to the actor's state
299 * @param clientActor A reference to the client who sent this message. This
300 * is the same reference that was passed to persistData
301 * by the derived actor. clientActor may be null when
302 * the RaftActor is behaving as a follower or during
304 * @param identifier The identifier of the persisted data. This is also
305 * the same identifier that was passed to persistData by
306 * the derived actor. identifier may be null when
307 * the RaftActor is behaving as a follower or during
309 * @param data A piece of data that was persisted by the persistData call.
310 * This should NEVER be null.
312 protected abstract void applyState(ActorRef clientActor, String identifier,
316 * This method will be called by the RaftActor when a snapshot needs to be
317 * created. The derived actor should respond with its current state.
319 * During recovery the state that is returned by the derived actor will
320 * be passed back to it by calling the applySnapshot method
322 * @return The current state of the actor
324 protected abstract Object createSnapshot();
327 * This method will be called by the RaftActor during recovery to
328 * reconstruct the state of the actor.
330 * This method may also be called at any other point during normal
331 * operations when the derived actor is out of sync with it's peers
332 * and the only way to bring it in sync is by applying a snapshot
334 * @param snapshot A snapshot of the state of the actor
336 protected abstract void applySnapshot(Object snapshot);
338 private RaftActorBehavior switchBehavior(RaftState state) {
339 if (currentBehavior != null) {
340 if (currentBehavior.state() == state) {
341 return currentBehavior;
343 LOG.info("Switching from state " + currentBehavior.state() + " to "
347 currentBehavior.close();
348 } catch (Exception e) {
350 "Failed to close behavior : " + currentBehavior.state());
354 LOG.info("Switching behavior to " + state);
356 RaftActorBehavior behavior = null;
357 if (state == RaftState.Candidate) {
358 behavior = new Candidate(context);
359 } else if (state == RaftState.Follower) {
360 behavior = new Follower(context);
362 behavior = new Leader(context);
367 private void trimPersistentData(long sequenceNumber) {
368 // Trim akka snapshots
369 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
370 // For now guessing that it is ANDed.
371 deleteSnapshots(new SnapshotSelectionCriteria(
372 sequenceNumber - SNAPSHOT_ENTRY_COUNT, 43200000));
375 deleteMessages(sequenceNumber);
379 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
381 public ReplicatedLogImpl(Snapshot snapshot) {
382 super(snapshot.getState(),
383 snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
384 snapshot.getUnAppliedEntries());
387 public ReplicatedLogImpl() {
391 @Override public void removeFromAndPersist(long logEntryIndex) {
392 int adjustedIndex = adjustedIndex(logEntryIndex);
394 if (adjustedIndex < 0) {
398 // FIXME: Maybe this should be done after the command is saved
399 journal.subList(adjustedIndex , journal.size()).clear();
401 persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
403 @Override public void apply(DeleteEntries param)
405 //FIXME : Doing nothing for now
410 @Override public void appendAndPersist(
411 final ReplicatedLogEntry replicatedLogEntry) {
412 appendAndPersist(null, null, replicatedLogEntry);
415 public void appendAndPersist(final ActorRef clientActor,
416 final String identifier,
417 final ReplicatedLogEntry replicatedLogEntry) {
418 context.getLogger().debug(
419 "Append log entry and persist {} ", replicatedLogEntry);
420 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
421 journal.add(replicatedLogEntry);
423 // When persisting events with persist it is guaranteed that the
424 // persistent actor will not receive further commands between the
425 // persist call and the execution(s) of the associated event
426 // handler. This also holds for multiple persist calls in context
427 // of a single command.
428 persist(replicatedLogEntry,
429 new Procedure<ReplicatedLogEntry>() {
430 public void apply(ReplicatedLogEntry evt) throws Exception {
431 // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned.
432 if (journal.size() > SNAPSHOT_ENTRY_COUNT) {
433 LOG.info("Initiating Snapshot Capture..");
434 long lastAppliedIndex = -1;
435 long lastAppliedTerm = -1;
437 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
438 if (lastAppliedEntry != null) {
439 lastAppliedIndex = lastAppliedEntry.getIndex();
440 lastAppliedTerm = lastAppliedEntry.getTerm();
443 LOG.debug("Snapshot Capture logSize: {}", journal.size());
444 LOG.debug("Snapshot Capture lastApplied:{} ", context.getLastApplied());
445 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
446 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
448 // create a snapshot object from the state provided and save it
449 // when snapshot is saved async, SaveSnapshotSuccess is raised.
450 Snapshot sn = Snapshot.create(createSnapshot(),
451 getFrom(context.getLastApplied() + 1),
452 lastIndex(), lastTerm(), lastAppliedIndex,
456 LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
458 //be greedy and remove entries from in-mem journal which are in the snapshot
459 // and update snapshotIndex and snapshotTerm without waiting for the success,
460 // TODO: damage-recovery to be done on failure
461 journal.subList(0, (int) (lastAppliedIndex - snapshotIndex)).clear();
462 snapshotIndex = lastAppliedIndex;
463 snapshotTerm = lastAppliedTerm;
465 LOG.info("Removed in-memory snapshotted entries, " +
466 "adjusted snaphsotIndex:{}" +
467 "and term:{}", snapshotIndex, lastAppliedTerm);
469 // Send message for replication
470 if (clientActor != null) {
471 currentBehavior.handleMessage(getSelf(),
472 new Replicate(clientActor, identifier,
483 private static class DeleteEntries implements Serializable {
484 private final int fromIndex;
487 public DeleteEntries(int fromIndex) {
488 this.fromIndex = fromIndex;
491 public int getFromIndex() {
497 private static class Snapshot implements Serializable {
498 private final Object state;
499 private final List<ReplicatedLogEntry> unAppliedEntries;
500 private final long lastIndex;
501 private final long lastTerm;
502 private final long lastAppliedIndex;
503 private final long lastAppliedTerm;
505 private Snapshot(Object state,
506 List<ReplicatedLogEntry> unAppliedEntries, long lastIndex,
507 long lastTerm, long lastAppliedIndex, long lastAppliedTerm) {
509 this.unAppliedEntries = unAppliedEntries;
510 this.lastIndex = lastIndex;
511 this.lastTerm = lastTerm;
512 this.lastAppliedIndex = lastAppliedIndex;
513 this.lastAppliedTerm = lastAppliedTerm;
517 public static Snapshot create(Object state,
518 List<ReplicatedLogEntry> entries, long lastIndex, long lastTerm,
519 long lastAppliedIndex, long lastAppliedTerm) {
520 return new Snapshot(state, entries, lastIndex, lastTerm,
521 lastAppliedIndex, lastAppliedTerm);
524 public Object getState() {
528 public List<ReplicatedLogEntry> getUnAppliedEntries() {
529 return unAppliedEntries;
532 public long getLastTerm() {
536 public long getLastAppliedIndex() {
537 return lastAppliedIndex;
540 public long getLastAppliedTerm() {
541 return lastAppliedTerm;
544 public String getLogMessage() {
545 StringBuilder sb = new StringBuilder();
546 return sb.append("Snapshot={")
547 .append("lastTerm:" + this.getLastTerm() + ", ")
548 .append("LastAppliedIndex:" + this.getLastAppliedIndex() + ", ")
549 .append("LastAppliedTerm:" + this.getLastAppliedTerm() + ", ")
550 .append("UnAppliedEntries size:" + this.getUnAppliedEntries().size() + "}")
556 private class ElectionTermImpl implements ElectionTerm {
558 * Identifier of the actor whose election term information this is
560 private long currentTerm = 0;
561 private String votedFor = null;
563 public long getCurrentTerm() {
567 public String getVotedFor() {
571 @Override public void update(long currentTerm, String votedFor) {
572 LOG.info("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
574 this.currentTerm = currentTerm;
575 this.votedFor = votedFor;
579 public void updateAndPersist(long currentTerm, String votedFor){
580 update(currentTerm, votedFor);
581 // FIXME : Maybe first persist then update the state
582 persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
584 @Override public void apply(UpdateElectionTerm param)
592 private static class UpdateElectionTerm implements Serializable {
593 private final long currentTerm;
594 private final String votedFor;
596 public UpdateElectionTerm(long currentTerm, String votedFor) {
597 this.currentTerm = currentTerm;
598 this.votedFor = votedFor;
601 public long getCurrentTerm() {
605 public String getVotedFor() {