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.japi.Procedure;
14 import akka.persistence.RecoveryCompleted;
15 import akka.persistence.SaveSnapshotFailure;
16 import akka.persistence.SaveSnapshotSuccess;
17 import akka.persistence.SnapshotOffer;
18 import akka.persistence.SnapshotSelectionCriteria;
19 import com.google.common.annotations.VisibleForTesting;
20 import com.google.common.base.Optional;
21 import com.google.common.base.Stopwatch;
22 import com.google.protobuf.ByteString;
23 import java.io.Serializable;
25 import org.opendaylight.controller.cluster.DataPersistenceProvider;
26 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
27 import org.opendaylight.controller.cluster.notifications.RoleChanged;
28 import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
29 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
30 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
31 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
32 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
33 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
34 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
35 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
36 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
37 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
38 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
39 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
40 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
41 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
42 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
43 import org.opendaylight.controller.protobuff.messages.cluster.raft.AppendEntriesMessages;
44 import org.slf4j.Logger;
45 import org.slf4j.LoggerFactory;
48 * RaftActor encapsulates a state machine that needs to be kept synchronized
49 * in a cluster. It implements the RAFT algorithm as described in the paper
50 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
51 * In Search of an Understandable Consensus Algorithm</a>
53 * RaftActor has 3 states and each state has a certain behavior associated
54 * with it. A Raft actor can behave as,
57 * <li> A Follower (or) </li>
58 * <li> A Candidate </li>
62 * A RaftActor MUST be a Leader in order to accept requests from clients to
63 * change the state of it's encapsulated state machine. Once a RaftActor becomes
64 * a Leader it is also responsible for ensuring that all followers ultimately
65 * have the same log and therefore the same state machine as itself.
68 * The current behavior of a RaftActor determines how election for leadership
69 * is initiated and how peer RaftActors react to request for votes.
72 * Each RaftActor also needs to know the current election term. It uses this
73 * information for a couple of things. One is to simply figure out who it
74 * voted for in the last election. Another is to figure out if the message
75 * it received to update it's state is stale.
78 * The RaftActor uses akka-persistence to store it's replicated log.
79 * Furthermore through it's behaviors a Raft Actor determines
82 * <li> when a log entry should be persisted </li>
83 * <li> when a log entry should be applied to the state machine (and) </li>
84 * <li> when a snapshot should be saved </li>
87 public abstract class RaftActor extends AbstractUntypedPersistentActor {
88 protected final Logger LOG = LoggerFactory.getLogger(getClass());
91 * The current state determines the current behavior of a RaftActor
92 * A Raft Actor always starts off in the Follower State
94 private RaftActorBehavior currentBehavior;
97 * This context should NOT be passed directly to any other actor it is
98 * only to be consumed by the RaftActorBehaviors
100 private final RaftActorContext context;
103 * The in-memory journal
105 private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
107 private CaptureSnapshot captureSnapshot = null;
109 private Stopwatch recoveryTimer;
111 private int currentRecoveryBatchCount;
113 public RaftActor(String id, Map<String, String> peerAddresses) {
114 this(id, peerAddresses, Optional.<ConfigParams>absent());
117 public RaftActor(String id, Map<String, String> peerAddresses,
118 Optional<ConfigParams> configParams) {
120 context = new RaftActorContextImpl(this.getSelf(),
121 this.getContext(), id, new ElectionTermImpl(),
122 -1, -1, replicatedLog, peerAddresses,
123 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
127 private void initRecoveryTimer() {
128 if(recoveryTimer == null) {
129 recoveryTimer = Stopwatch.createStarted();
134 public void preStart() throws Exception {
135 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
136 context.getConfigParams().getJournalRecoveryLogBatchSize());
142 public void handleRecover(Object message) {
143 if(persistence().isRecoveryApplicable()) {
144 if (message instanceof SnapshotOffer) {
145 onRecoveredSnapshot((SnapshotOffer) message);
146 } else if (message instanceof ReplicatedLogEntry) {
147 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
148 } else if (message instanceof ApplyLogEntries) {
149 onRecoveredApplyLogEntries((ApplyLogEntries) message);
150 } else if (message instanceof DeleteEntries) {
151 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
152 } else if (message instanceof UpdateElectionTerm) {
153 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
154 ((UpdateElectionTerm) message).getVotedFor());
155 } else if (message instanceof RecoveryCompleted) {
156 onRecoveryCompletedMessage();
159 if (message instanceof RecoveryCompleted) {
160 // Delete all the messages from the akka journal so that we do not end up with consistency issues
161 // Note I am not using the dataPersistenceProvider and directly using the akka api here
162 deleteMessages(lastSequenceNr());
164 // Delete all the akka snapshots as they will not be needed
165 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
167 onRecoveryComplete();
169 initializeBehavior();
174 private void onRecoveredSnapshot(SnapshotOffer offer) {
175 if(LOG.isDebugEnabled()) {
176 LOG.debug("{}: SnapshotOffer called..", persistenceId());
181 Snapshot snapshot = (Snapshot) offer.snapshot();
183 // Create a replicated log with the snapshot information
184 // The replicated log can be used later on to retrieve this snapshot
185 // when we need to install it on a peer
186 replicatedLog = new ReplicatedLogImpl(snapshot);
188 context.setReplicatedLog(replicatedLog);
189 context.setLastApplied(snapshot.getLastAppliedIndex());
190 context.setCommitIndex(snapshot.getLastAppliedIndex());
192 Stopwatch timer = Stopwatch.createStarted();
194 // Apply the snapshot to the actors state
195 applyRecoverySnapshot(snapshot.getState());
198 LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
199 replicatedLog.size(), persistenceId(), timer.toString(),
200 replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm());
203 private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
204 if(LOG.isDebugEnabled()) {
205 LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex());
208 replicatedLog.append(logEntry);
211 private void onRecoveredApplyLogEntries(ApplyLogEntries ale) {
212 if(LOG.isDebugEnabled()) {
213 LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}",
214 persistenceId(), context.getLastApplied() + 1, ale.getToIndex());
217 for (long i = context.getLastApplied() + 1; i <= ale.getToIndex(); i++) {
218 batchRecoveredLogEntry(replicatedLog.get(i));
221 context.setLastApplied(ale.getToIndex());
222 context.setCommitIndex(ale.getToIndex());
225 private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
228 int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
229 if(currentRecoveryBatchCount == 0) {
230 startLogRecoveryBatch(batchSize);
233 appendRecoveredLogEntry(logEntry.getData());
235 if(++currentRecoveryBatchCount >= batchSize) {
236 endCurrentLogRecoveryBatch();
240 private void endCurrentLogRecoveryBatch() {
241 applyCurrentLogRecoveryBatch();
242 currentRecoveryBatchCount = 0;
245 private void onRecoveryCompletedMessage() {
246 if(currentRecoveryBatchCount > 0) {
247 endCurrentLogRecoveryBatch();
250 onRecoveryComplete();
252 String recoveryTime = "";
253 if(recoveryTimer != null) {
254 recoveryTimer.stop();
255 recoveryTime = " in " + recoveryTimer.toString();
256 recoveryTimer = null;
260 "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
261 "Persistence Id = " + persistenceId() +
262 " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
264 replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(),
265 replicatedLog.getSnapshotTerm(), replicatedLog.size());
267 initializeBehavior();
270 protected void initializeBehavior(){
271 changeCurrentBehavior(new Follower(context));
274 protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
275 RaftActorBehavior oldBehavior = currentBehavior;
276 currentBehavior = newBehavior;
277 handleBehaviorChange(oldBehavior, currentBehavior);
280 @Override public void handleCommand(Object message) {
281 if (message instanceof ApplyState){
282 ApplyState applyState = (ApplyState) message;
284 if(LOG.isDebugEnabled()) {
285 LOG.debug("{}: Applying state for log index {} data {}",
286 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
287 applyState.getReplicatedLogEntry().getData());
290 applyState(applyState.getClientActor(), applyState.getIdentifier(),
291 applyState.getReplicatedLogEntry().getData());
293 } else if (message instanceof ApplyLogEntries){
294 ApplyLogEntries ale = (ApplyLogEntries) message;
295 if(LOG.isDebugEnabled()) {
296 LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), ale.getToIndex());
298 persistence().persist(new ApplyLogEntries(ale.getToIndex()), new Procedure<ApplyLogEntries>() {
300 public void apply(ApplyLogEntries param) throws Exception {
304 } else if(message instanceof ApplySnapshot ) {
305 Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
307 if(LOG.isDebugEnabled()) {
308 LOG.debug("{}: ApplySnapshot called on Follower Actor " +
309 "snapshotIndex:{}, snapshotTerm:{}", persistenceId(), snapshot.getLastAppliedIndex(),
310 snapshot.getLastAppliedTerm()
314 applySnapshot(snapshot.getState());
316 //clears the followers log, sets the snapshot index to ensure adjusted-index works
317 replicatedLog = new ReplicatedLogImpl(snapshot);
318 context.setReplicatedLog(replicatedLog);
319 context.setLastApplied(snapshot.getLastAppliedIndex());
321 } else if (message instanceof FindLeader) {
323 new FindLeaderReply(getLeaderAddress()),
327 } else if (message instanceof SaveSnapshotSuccess) {
328 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
329 LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId());
331 long sequenceNumber = success.metadata().sequenceNr();
333 commitSnapshot(sequenceNumber);
335 } else if (message instanceof SaveSnapshotFailure) {
336 SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
338 LOG.error("{}: SaveSnapshotFailure received for snapshot Cause:",
339 persistenceId(), saveSnapshotFailure.cause());
341 context.getReplicatedLog().snapshotRollback();
343 LOG.info("{}: Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
344 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}", persistenceId(),
345 context.getReplicatedLog().getSnapshotIndex(),
346 context.getReplicatedLog().getSnapshotTerm(),
347 context.getReplicatedLog().size());
349 } else if (message instanceof CaptureSnapshot) {
350 LOG.info("{}: CaptureSnapshot received by actor", persistenceId());
352 if(captureSnapshot == null) {
353 captureSnapshot = (CaptureSnapshot)message;
357 } else if (message instanceof CaptureSnapshotReply){
358 handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot());
361 if (!(message instanceof AppendEntriesMessages.AppendEntries)
362 && !(message instanceof AppendEntriesReply) && !(message instanceof SendHeartBeat)) {
363 if(LOG.isDebugEnabled()) {
364 LOG.debug("{}: onReceiveCommand: message: {}", persistenceId(), message.getClass());
368 RaftActorBehavior oldBehavior = currentBehavior;
369 currentBehavior = currentBehavior.handleMessage(getSender(), message);
371 handleBehaviorChange(oldBehavior, currentBehavior);
375 private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) {
376 if (oldBehavior != currentBehavior){
380 String oldBehaviorLeaderId = oldBehavior == null? null : oldBehavior.getLeaderId();
381 String oldBehaviorState = oldBehavior == null? null : oldBehavior.state().name();
383 // it can happen that the state has not changed but the leader has changed.
384 onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
386 if (getRoleChangeNotifier().isPresent() &&
387 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
388 getRoleChangeNotifier().get().tell(
389 new RoleChanged(getId(), oldBehaviorState , currentBehavior.state().name()),
395 * When a derived RaftActor needs to persist something it must call
402 protected void persistData(final ActorRef clientActor, final String identifier,
403 final Payload data) {
405 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
406 context.getReplicatedLog().lastIndex() + 1,
407 context.getTermInformation().getCurrentTerm(), data);
409 if(LOG.isDebugEnabled()) {
410 LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
413 final RaftActorContext raftContext = getRaftActorContext();
416 .appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
418 public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
420 // Increment the Commit Index and the Last Applied values
421 raftContext.setCommitIndex(replicatedLogEntry.getIndex());
422 raftContext.setLastApplied(replicatedLogEntry.getIndex());
424 // Apply the state immediately
425 applyState(clientActor, identifier, data);
427 // Send a ApplyLogEntries message so that we write the fact that we applied
428 // the state to durable storage
429 self().tell(new ApplyLogEntries((int) replicatedLogEntry.getIndex()), self());
431 // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot
432 if(!context.isSnapshotCaptureInitiated()){
433 raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(),
434 raftContext.getTermInformation().getCurrentTerm());
435 raftContext.getReplicatedLog().snapshotCommit();
437 LOG.debug("{}: Skipping fake snapshotting for {} because real snapshotting is in progress",
438 persistenceId(), getId());
440 } else if (clientActor != null) {
441 // Send message for replication
442 currentBehavior.handleMessage(getSelf(),
443 new Replicate(clientActor, identifier,
451 protected String getId() {
452 return context.getId();
456 * Derived actors can call the isLeader method to check if the current
457 * RaftActor is the Leader or not
459 * @return true it this RaftActor is a Leader false otherwise
461 protected boolean isLeader() {
462 return context.getId().equals(currentBehavior.getLeaderId());
466 * Derived actor can call getLeader if they need a reference to the Leader.
467 * This would be useful for example in forwarding a request to an actor
468 * which is the leader
470 * @return A reference to the leader if known, null otherwise
472 protected ActorSelection getLeader(){
473 String leaderAddress = getLeaderAddress();
475 if(leaderAddress == null){
479 return context.actorSelection(leaderAddress);
484 * @return the current leader's id
486 protected String getLeaderId(){
487 return currentBehavior.getLeaderId();
490 protected RaftState getRaftState() {
491 return currentBehavior.state();
494 protected ReplicatedLogEntry getLastLogEntry() {
495 return replicatedLog.last();
498 protected Long getCurrentTerm(){
499 return context.getTermInformation().getCurrentTerm();
502 protected Long getCommitIndex(){
503 return context.getCommitIndex();
506 protected Long getLastApplied(){
507 return context.getLastApplied();
510 protected RaftActorContext getRaftActorContext() {
515 * setPeerAddress sets the address of a known peer at a later time.
517 * This is to account for situations where a we know that a peer
518 * exists but we do not know an address up-front. This may also be used in
519 * situations where a known peer starts off in a different location and we
520 * need to change it's address
522 * Note that if the peerId does not match the list of peers passed to
523 * this actor during construction an IllegalStateException will be thrown.
528 protected void setPeerAddress(String peerId, String peerAddress){
529 context.setPeerAddress(peerId, peerAddress);
532 protected void commitSnapshot(long sequenceNumber) {
533 context.getReplicatedLog().snapshotCommit();
535 // TODO: Not sure if we want to be this aggressive with trimming stuff
536 trimPersistentData(sequenceNumber);
540 * The applyState method will be called by the RaftActor when some data
541 * needs to be applied to the actor's state
543 * @param clientActor A reference to the client who sent this message. This
544 * is the same reference that was passed to persistData
545 * by the derived actor. clientActor may be null when
546 * the RaftActor is behaving as a follower or during
548 * @param identifier The identifier of the persisted data. This is also
549 * the same identifier that was passed to persistData by
550 * the derived actor. identifier may be null when
551 * the RaftActor is behaving as a follower or during
553 * @param data A piece of data that was persisted by the persistData call.
554 * This should NEVER be null.
556 protected abstract void applyState(ActorRef clientActor, String identifier,
560 * This method is called during recovery at the start of a batch of state entries. Derived
561 * classes should perform any initialization needed to start a batch.
563 protected abstract void startLogRecoveryBatch(int maxBatchSize);
566 * This method is called during recovery to append state data to the current batch. This method
567 * is called 1 or more times after {@link #startLogRecoveryBatch}.
569 * @param data the state data
571 protected abstract void appendRecoveredLogEntry(Payload data);
574 * This method is called during recovery to reconstruct the state of the actor.
576 * @param snapshot A snapshot of the state of the actor
578 protected abstract void applyRecoverySnapshot(byte[] snapshotBytes);
581 * This method is called during recovery at the end of a batch to apply the current batched
582 * log entries. This method is called after {@link #appendRecoveredLogEntry}.
584 protected abstract void applyCurrentLogRecoveryBatch();
587 * This method is called when recovery is complete.
589 protected abstract void onRecoveryComplete();
592 * This method will be called by the RaftActor when a snapshot needs to be
593 * created. The derived actor should respond with its current state.
595 * During recovery the state that is returned by the derived actor will
596 * be passed back to it by calling the applySnapshot method
598 * @return The current state of the actor
600 protected abstract void createSnapshot();
603 * This method can be called at any other point during normal
604 * operations when the derived actor is out of sync with it's peers
605 * and the only way to bring it in sync is by applying a snapshot
607 * @param snapshotBytes A snapshot of the state of the actor
609 protected abstract void applySnapshot(byte[] snapshotBytes);
612 * This method will be called by the RaftActor when the state of the
613 * RaftActor changes. The derived actor can then use methods like
614 * isLeader or getLeader to do something useful
616 protected abstract void onStateChanged();
618 protected abstract DataPersistenceProvider persistence();
621 * Notifier Actor for this RaftActor to notify when a role change happens
622 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
624 protected abstract Optional<ActorRef> getRoleChangeNotifier();
626 protected void onLeaderChanged(String oldLeader, String newLeader){};
628 private void trimPersistentData(long sequenceNumber) {
629 // Trim akka snapshots
630 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
631 // For now guessing that it is ANDed.
632 persistence().deleteSnapshots(new SnapshotSelectionCriteria(
633 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
636 persistence().deleteMessages(sequenceNumber);
639 private String getLeaderAddress(){
641 return getSelf().path().toString();
643 String leaderId = currentBehavior.getLeaderId();
644 if (leaderId == null) {
647 String peerAddress = context.getPeerAddress(leaderId);
648 if(LOG.isDebugEnabled()) {
649 LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
650 persistenceId(), leaderId, peerAddress);
656 private void handleCaptureSnapshotReply(byte[] snapshotBytes) {
657 LOG.info("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length);
659 // create a snapshot object from the state provided and save it
660 // when snapshot is saved async, SaveSnapshotSuccess is raised.
662 Snapshot sn = Snapshot.create(snapshotBytes,
663 context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
664 captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
665 captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
667 persistence().saveSnapshot(sn);
669 LOG.info("{}: Persisting of snapshot done:{}", persistenceId(), sn.getLogMessage());
671 //be greedy and remove entries from in-mem journal which are in the snapshot
672 // and update snapshotIndex and snapshotTerm without waiting for the success,
674 context.getReplicatedLog().snapshotPreCommit(
675 captureSnapshot.getLastAppliedIndex(),
676 captureSnapshot.getLastAppliedTerm());
678 LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
679 "and term:{}", persistenceId(), captureSnapshot.getLastAppliedIndex(),
680 captureSnapshot.getLastAppliedTerm());
682 if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
683 // this would be call straight to the leader and won't initiate in serialization
684 currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(
685 ByteString.copyFrom(snapshotBytes)));
688 captureSnapshot = null;
689 context.setSnapshotCaptureInitiated(false);
692 protected boolean hasFollowers(){
693 return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
696 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
698 private static final int DATA_SIZE_DIVIDER = 5;
699 private long dataSizeSinceLastSnapshot = 0;
701 public ReplicatedLogImpl(Snapshot snapshot) {
702 super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
703 snapshot.getUnAppliedEntries());
706 public ReplicatedLogImpl() {
710 @Override public void removeFromAndPersist(long logEntryIndex) {
711 int adjustedIndex = adjustedIndex(logEntryIndex);
713 if (adjustedIndex < 0) {
717 // FIXME: Maybe this should be done after the command is saved
718 journal.subList(adjustedIndex , journal.size()).clear();
720 persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
722 @Override public void apply(DeleteEntries param)
724 //FIXME : Doing nothing for now
726 for(ReplicatedLogEntry entry : journal){
727 dataSize += entry.size();
733 @Override public void appendAndPersist(
734 final ReplicatedLogEntry replicatedLogEntry) {
735 appendAndPersist(replicatedLogEntry, null);
739 public int dataSize() {
743 public void appendAndPersist(
744 final ReplicatedLogEntry replicatedLogEntry,
745 final Procedure<ReplicatedLogEntry> callback) {
747 if(LOG.isDebugEnabled()) {
748 LOG.debug("{}: Append log entry and persist {} ", persistenceId(), replicatedLogEntry);
751 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
752 journal.add(replicatedLogEntry);
754 // When persisting events with persist it is guaranteed that the
755 // persistent actor will not receive further commands between the
756 // persist call and the execution(s) of the associated event
757 // handler. This also holds for multiple persist calls in context
758 // of a single command.
759 persistence().persist(replicatedLogEntry,
760 new Procedure<ReplicatedLogEntry>() {
762 public void apply(ReplicatedLogEntry evt) throws Exception {
763 int logEntrySize = replicatedLogEntry.size();
765 dataSize += logEntrySize;
766 long dataSizeForCheck = dataSize;
768 dataSizeSinceLastSnapshot += logEntrySize;
769 long journalSize = lastIndex()+1;
771 if(!hasFollowers()) {
772 // When we do not have followers we do not maintain an in-memory log
773 // due to this the journalSize will never become anything close to the
774 // snapshot batch count. In fact will mostly be 1.
775 // Similarly since the journal's dataSize depends on the entries in the
776 // journal the journal's dataSize will never reach a value close to the
778 // By maintaining the dataSize outside the journal we are tracking essentially
779 // what we have written to the disk however since we no longer are in
780 // need of doing a snapshot just for the sake of freeing up memory we adjust
781 // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
782 // as if we were maintaining a real snapshot
783 dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
786 long dataThreshold = Runtime.getRuntime().totalMemory() *
787 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
789 // when a snaphsot is being taken, captureSnapshot != null
790 if (!context.isSnapshotCaptureInitiated() &&
791 ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
792 dataSizeForCheck > dataThreshold)) {
794 dataSizeSinceLastSnapshot = 0;
796 LOG.info("{}: Initiating Snapshot Capture..", persistenceId());
797 long lastAppliedIndex = -1;
798 long lastAppliedTerm = -1;
800 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
801 if (!hasFollowers()) {
802 lastAppliedIndex = replicatedLogEntry.getIndex();
803 lastAppliedTerm = replicatedLogEntry.getTerm();
804 } else if (lastAppliedEntry != null) {
805 lastAppliedIndex = lastAppliedEntry.getIndex();
806 lastAppliedTerm = lastAppliedEntry.getTerm();
809 if(LOG.isDebugEnabled()) {
810 LOG.debug("{}: Snapshot Capture logSize: {}", persistenceId(), journal.size());
811 LOG.debug("{}: Snapshot Capture lastApplied:{} ",
812 persistenceId(), context.getLastApplied());
813 LOG.debug("{}: Snapshot Capture lastAppliedIndex:{}", persistenceId(),
815 LOG.debug("{}: Snapshot Capture lastAppliedTerm:{}", persistenceId(),
819 // send a CaptureSnapshot to self to make the expensive operation async.
820 getSelf().tell(new CaptureSnapshot(
821 lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm),
823 context.setSnapshotCaptureInitiated(true);
825 if(callback != null){
826 callback.apply(replicatedLogEntry);
835 static class DeleteEntries implements Serializable {
836 private static final long serialVersionUID = 1L;
837 private final int fromIndex;
839 public DeleteEntries(int fromIndex) {
840 this.fromIndex = fromIndex;
843 public int getFromIndex() {
849 private class ElectionTermImpl implements ElectionTerm {
851 * Identifier of the actor whose election term information this is
853 private long currentTerm = 0;
854 private String votedFor = null;
857 public long getCurrentTerm() {
862 public String getVotedFor() {
866 @Override public void update(long currentTerm, String votedFor) {
867 if(LOG.isDebugEnabled()) {
868 LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
870 this.currentTerm = currentTerm;
871 this.votedFor = votedFor;
875 public void updateAndPersist(long currentTerm, String votedFor){
876 update(currentTerm, votedFor);
877 // FIXME : Maybe first persist then update the state
878 persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
880 @Override public void apply(UpdateElectionTerm param)
888 static class UpdateElectionTerm implements Serializable {
889 private static final long serialVersionUID = 1L;
890 private final long currentTerm;
891 private final String votedFor;
893 public UpdateElectionTerm(long currentTerm, String votedFor) {
894 this.currentTerm = currentTerm;
895 this.votedFor = votedFor;
898 public long getCurrentTerm() {
902 public String getVotedFor() {
907 protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
909 public NonPersistentRaftDataProvider(){
914 * The way snapshotting works is,
916 * <li> RaftActor calls createSnapshot on the Shard
917 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
918 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
919 * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
920 * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
921 * in SaveSnapshotSuccess.
926 public void saveSnapshot(Object o) {
927 // Make saving Snapshot successful
933 void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
934 currentBehavior = behavior;
937 protected RaftActorBehavior getCurrentBehavior() {
938 return currentBehavior;