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 java.util.concurrent.TimeUnit;
26 import org.opendaylight.controller.cluster.DataPersistenceProvider;
27 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
28 import org.opendaylight.controller.cluster.notifications.RoleChanged;
29 import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
30 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
31 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
32 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
33 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
34 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
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.protobuff.client.messages.Payload;
42 import org.slf4j.Logger;
43 import org.slf4j.LoggerFactory;
46 * RaftActor encapsulates a state machine that needs to be kept synchronized
47 * in a cluster. It implements the RAFT algorithm as described in the paper
48 * <a href='https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf'>
49 * In Search of an Understandable Consensus Algorithm</a>
51 * RaftActor has 3 states and each state has a certain behavior associated
52 * with it. A Raft actor can behave as,
55 * <li> A Follower (or) </li>
56 * <li> A Candidate </li>
60 * A RaftActor MUST be a Leader in order to accept requests from clients to
61 * change the state of it's encapsulated state machine. Once a RaftActor becomes
62 * a Leader it is also responsible for ensuring that all followers ultimately
63 * have the same log and therefore the same state machine as itself.
66 * The current behavior of a RaftActor determines how election for leadership
67 * is initiated and how peer RaftActors react to request for votes.
70 * Each RaftActor also needs to know the current election term. It uses this
71 * information for a couple of things. One is to simply figure out who it
72 * voted for in the last election. Another is to figure out if the message
73 * it received to update it's state is stale.
76 * The RaftActor uses akka-persistence to store it's replicated log.
77 * Furthermore through it's behaviors a Raft Actor determines
80 * <li> when a log entry should be persisted </li>
81 * <li> when a log entry should be applied to the state machine (and) </li>
82 * <li> when a snapshot should be saved </li>
85 public abstract class RaftActor extends AbstractUntypedPersistentActor {
87 private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
89 protected final Logger LOG = LoggerFactory.getLogger(getClass());
92 * The current state determines the current behavior of a RaftActor
93 * A Raft Actor always starts off in the Follower State
95 private RaftActorBehavior currentBehavior;
98 * This context should NOT be passed directly to any other actor it is
99 * only to be consumed by the RaftActorBehaviors
101 private final RaftActorContext context;
104 * The in-memory journal
106 private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
108 private CaptureSnapshot captureSnapshot = null;
110 private Stopwatch recoveryTimer;
112 private int currentRecoveryBatchCount;
114 public RaftActor(String id, Map<String, String> peerAddresses) {
115 this(id, peerAddresses, Optional.<ConfigParams>absent());
118 public RaftActor(String id, Map<String, String> peerAddresses,
119 Optional<ConfigParams> configParams) {
121 context = new RaftActorContextImpl(this.getSelf(),
122 this.getContext(), id, new ElectionTermImpl(),
123 -1, -1, replicatedLog, peerAddresses,
124 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
128 private void initRecoveryTimer() {
129 if(recoveryTimer == null) {
130 recoveryTimer = Stopwatch.createStarted();
135 public void preStart() throws Exception {
136 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
137 context.getConfigParams().getJournalRecoveryLogBatchSize());
143 public void handleRecover(Object message) {
144 if(persistence().isRecoveryApplicable()) {
145 if (message instanceof SnapshotOffer) {
146 onRecoveredSnapshot((SnapshotOffer) message);
147 } else if (message instanceof ReplicatedLogEntry) {
148 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
149 } else if (message instanceof ApplyLogEntries) {
150 onRecoveredApplyLogEntries((ApplyLogEntries) message);
151 } else if (message instanceof DeleteEntries) {
152 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
153 } else if (message instanceof UpdateElectionTerm) {
154 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
155 ((UpdateElectionTerm) message).getVotedFor());
156 } else if (message instanceof RecoveryCompleted) {
157 onRecoveryCompletedMessage();
160 if (message instanceof RecoveryCompleted) {
161 // Delete all the messages from the akka journal so that we do not end up with consistency issues
162 // Note I am not using the dataPersistenceProvider and directly using the akka api here
163 deleteMessages(lastSequenceNr());
165 // Delete all the akka snapshots as they will not be needed
166 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
168 onRecoveryComplete();
170 initializeBehavior();
175 private void onRecoveredSnapshot(SnapshotOffer offer) {
176 if(LOG.isDebugEnabled()) {
177 LOG.debug("{}: SnapshotOffer called..", persistenceId());
182 Snapshot snapshot = (Snapshot) offer.snapshot();
184 // Create a replicated log with the snapshot information
185 // The replicated log can be used later on to retrieve this snapshot
186 // when we need to install it on a peer
187 replicatedLog = new ReplicatedLogImpl(snapshot);
189 context.setReplicatedLog(replicatedLog);
190 context.setLastApplied(snapshot.getLastAppliedIndex());
191 context.setCommitIndex(snapshot.getLastAppliedIndex());
193 Stopwatch timer = Stopwatch.createStarted();
195 // Apply the snapshot to the actors state
196 applyRecoverySnapshot(snapshot.getState());
199 LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
200 replicatedLog.size(), persistenceId(), timer.toString(),
201 replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm());
204 private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
205 if(LOG.isDebugEnabled()) {
206 LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex());
209 replicatedLog.append(logEntry);
212 private void onRecoveredApplyLogEntries(ApplyLogEntries ale) {
213 if(LOG.isDebugEnabled()) {
214 LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}",
215 persistenceId(), context.getLastApplied() + 1, ale.getToIndex());
218 for (long i = context.getLastApplied() + 1; i <= ale.getToIndex(); i++) {
219 batchRecoveredLogEntry(replicatedLog.get(i));
222 context.setLastApplied(ale.getToIndex());
223 context.setCommitIndex(ale.getToIndex());
226 private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
229 int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
230 if(currentRecoveryBatchCount == 0) {
231 startLogRecoveryBatch(batchSize);
234 appendRecoveredLogEntry(logEntry.getData());
236 if(++currentRecoveryBatchCount >= batchSize) {
237 endCurrentLogRecoveryBatch();
241 private void endCurrentLogRecoveryBatch() {
242 applyCurrentLogRecoveryBatch();
243 currentRecoveryBatchCount = 0;
246 private void onRecoveryCompletedMessage() {
247 if(currentRecoveryBatchCount > 0) {
248 endCurrentLogRecoveryBatch();
251 onRecoveryComplete();
253 String recoveryTime = "";
254 if(recoveryTimer != null) {
255 recoveryTimer.stop();
256 recoveryTime = " in " + recoveryTimer.toString();
257 recoveryTimer = null;
261 "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
262 "Persistence Id = " + persistenceId() +
263 " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
265 replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(),
266 replicatedLog.getSnapshotTerm(), replicatedLog.size());
268 initializeBehavior();
271 protected void initializeBehavior(){
272 changeCurrentBehavior(new Follower(context));
275 protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
276 RaftActorBehavior oldBehavior = currentBehavior;
277 currentBehavior = newBehavior;
278 handleBehaviorChange(oldBehavior, currentBehavior);
281 @Override public void handleCommand(Object message) {
282 if (message instanceof ApplyState){
283 ApplyState applyState = (ApplyState) message;
285 long elapsedTime = (System.nanoTime() - applyState.getStartTime());
286 if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
287 LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
288 TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
291 if(LOG.isDebugEnabled()) {
292 LOG.debug("{}: Applying state for log index {} data {}",
293 persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
294 applyState.getReplicatedLogEntry().getData());
297 applyState(applyState.getClientActor(), applyState.getIdentifier(),
298 applyState.getReplicatedLogEntry().getData());
300 } else if (message instanceof ApplyLogEntries){
301 ApplyLogEntries ale = (ApplyLogEntries) message;
302 if(LOG.isDebugEnabled()) {
303 LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), ale.getToIndex());
305 persistence().persist(new ApplyLogEntries(ale.getToIndex()), new Procedure<ApplyLogEntries>() {
307 public void apply(ApplyLogEntries param) throws Exception {
311 } else if(message instanceof ApplySnapshot ) {
312 Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
314 if(LOG.isDebugEnabled()) {
315 LOG.debug("{}: ApplySnapshot called on Follower Actor " +
316 "snapshotIndex:{}, snapshotTerm:{}", persistenceId(), snapshot.getLastAppliedIndex(),
317 snapshot.getLastAppliedTerm()
321 applySnapshot(snapshot.getState());
323 //clears the followers log, sets the snapshot index to ensure adjusted-index works
324 replicatedLog = new ReplicatedLogImpl(snapshot);
325 context.setReplicatedLog(replicatedLog);
326 context.setLastApplied(snapshot.getLastAppliedIndex());
328 } else if (message instanceof FindLeader) {
330 new FindLeaderReply(getLeaderAddress()),
334 } else if (message instanceof SaveSnapshotSuccess) {
335 SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
336 LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId());
338 long sequenceNumber = success.metadata().sequenceNr();
340 commitSnapshot(sequenceNumber);
342 } else if (message instanceof SaveSnapshotFailure) {
343 SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
345 LOG.error("{}: SaveSnapshotFailure received for snapshot Cause:",
346 persistenceId(), saveSnapshotFailure.cause());
348 context.getReplicatedLog().snapshotRollback();
350 LOG.info("{}: Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
351 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}", persistenceId(),
352 context.getReplicatedLog().getSnapshotIndex(),
353 context.getReplicatedLog().getSnapshotTerm(),
354 context.getReplicatedLog().size());
356 } else if (message instanceof CaptureSnapshot) {
357 LOG.info("{}: CaptureSnapshot received by actor", persistenceId());
359 if(captureSnapshot == null) {
360 captureSnapshot = (CaptureSnapshot)message;
364 } else if (message instanceof CaptureSnapshotReply){
365 handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot());
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 snapshotBytes 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 long dataThreshold = Runtime.getRuntime().totalMemory() *
672 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
673 if (context.getReplicatedLog().dataSize() > dataThreshold) {
674 // if memory is less, clear the log based on lastApplied.
675 // this could/should only happen if one of the followers is down
676 // as normally we keep removing from the log when its replicated to all.
677 context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getLastAppliedIndex(),
678 captureSnapshot.getLastAppliedTerm());
681 // clear the log based on replicatedToAllIndex
682 context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getReplicatedToAllIndex(),
683 captureSnapshot.getReplicatedToAllTerm());
685 getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
687 LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
688 "and term:{}", persistenceId(), captureSnapshot.getLastAppliedIndex(),
689 captureSnapshot.getLastAppliedTerm());
691 if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
692 // this would be call straight to the leader and won't initiate in serialization
693 currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(
694 ByteString.copyFrom(snapshotBytes)));
697 captureSnapshot = null;
698 context.setSnapshotCaptureInitiated(false);
701 protected boolean hasFollowers(){
702 return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
705 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
707 private static final int DATA_SIZE_DIVIDER = 5;
708 private long dataSizeSinceLastSnapshot = 0;
710 public ReplicatedLogImpl(Snapshot snapshot) {
711 super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
712 snapshot.getUnAppliedEntries());
715 public ReplicatedLogImpl() {
719 @Override public void removeFromAndPersist(long logEntryIndex) {
720 int adjustedIndex = adjustedIndex(logEntryIndex);
722 if (adjustedIndex < 0) {
726 // FIXME: Maybe this should be done after the command is saved
727 journal.subList(adjustedIndex , journal.size()).clear();
729 persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>() {
732 public void apply(DeleteEntries param)
734 //FIXME : Doing nothing for now
736 for (ReplicatedLogEntry entry : journal) {
737 dataSize += entry.size();
743 @Override public void appendAndPersist(
744 final ReplicatedLogEntry replicatedLogEntry) {
745 appendAndPersist(replicatedLogEntry, null);
748 public void appendAndPersist(
749 final ReplicatedLogEntry replicatedLogEntry,
750 final Procedure<ReplicatedLogEntry> callback) {
752 if(LOG.isDebugEnabled()) {
753 LOG.debug("{}: Append log entry and persist {} ", persistenceId(), replicatedLogEntry);
756 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
757 journal.add(replicatedLogEntry);
759 // When persisting events with persist it is guaranteed that the
760 // persistent actor will not receive further commands between the
761 // persist call and the execution(s) of the associated event
762 // handler. This also holds for multiple persist calls in context
763 // of a single command.
764 persistence().persist(replicatedLogEntry,
765 new Procedure<ReplicatedLogEntry>() {
767 public void apply(ReplicatedLogEntry evt) throws Exception {
768 int logEntrySize = replicatedLogEntry.size();
770 dataSize += logEntrySize;
771 long dataSizeForCheck = dataSize;
773 dataSizeSinceLastSnapshot += logEntrySize;
774 long journalSize = lastIndex() + 1;
776 if(!hasFollowers()) {
777 // When we do not have followers we do not maintain an in-memory log
778 // due to this the journalSize will never become anything close to the
779 // snapshot batch count. In fact will mostly be 1.
780 // Similarly since the journal's dataSize depends on the entries in the
781 // journal the journal's dataSize will never reach a value close to the
783 // By maintaining the dataSize outside the journal we are tracking essentially
784 // what we have written to the disk however since we no longer are in
785 // need of doing a snapshot just for the sake of freeing up memory we adjust
786 // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
787 // as if we were maintaining a real snapshot
788 dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
791 long dataThreshold = Runtime.getRuntime().totalMemory() *
792 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
794 // when a snaphsot is being taken, captureSnapshot != null
795 if (!context.isSnapshotCaptureInitiated() &&
796 ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
797 dataSizeForCheck > dataThreshold)) {
799 dataSizeSinceLastSnapshot = 0;
801 LOG.info("{}: Initiating Snapshot Capture, journalSize = {}, dataSizeForCheck = {}," +
802 " dataThreshold = {}", persistenceId(), journalSize, dataSizeForCheck, dataThreshold);
804 long lastAppliedIndex = -1;
805 long lastAppliedTerm = -1;
807 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
808 if (!hasFollowers()) {
809 lastAppliedIndex = replicatedLogEntry.getIndex();
810 lastAppliedTerm = replicatedLogEntry.getTerm();
811 } else if (lastAppliedEntry != null) {
812 lastAppliedIndex = lastAppliedEntry.getIndex();
813 lastAppliedTerm = lastAppliedEntry.getTerm();
816 if(LOG.isDebugEnabled()) {
817 LOG.debug("{}: Snapshot Capture logSize: {}", persistenceId(), journal.size());
818 LOG.debug("{}: Snapshot Capture lastApplied:{} ",
819 persistenceId(), context.getLastApplied());
820 LOG.debug("{}: Snapshot Capture lastAppliedIndex:{}", persistenceId(),
822 LOG.debug("{}: Snapshot Capture lastAppliedTerm:{}", persistenceId(),
826 // send a CaptureSnapshot to self to make the expensive operation async.
827 long replicatedToAllIndex = getCurrentBehavior().getReplicatedToAllIndex();
828 ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
829 getSelf().tell(new CaptureSnapshot(lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
830 (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
831 (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1)),
833 context.setSnapshotCaptureInitiated(true);
835 if (callback != null){
836 callback.apply(replicatedLogEntry);
845 static class DeleteEntries implements Serializable {
846 private static final long serialVersionUID = 1L;
847 private final int fromIndex;
849 public DeleteEntries(int fromIndex) {
850 this.fromIndex = fromIndex;
853 public int getFromIndex() {
859 private class ElectionTermImpl implements ElectionTerm {
861 * Identifier of the actor whose election term information this is
863 private long currentTerm = 0;
864 private String votedFor = null;
867 public long getCurrentTerm() {
872 public String getVotedFor() {
876 @Override public void update(long currentTerm, String votedFor) {
877 if(LOG.isDebugEnabled()) {
878 LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
880 this.currentTerm = currentTerm;
881 this.votedFor = votedFor;
885 public void updateAndPersist(long currentTerm, String votedFor){
886 update(currentTerm, votedFor);
887 // FIXME : Maybe first persist then update the state
888 persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
890 @Override public void apply(UpdateElectionTerm param)
898 static class UpdateElectionTerm implements Serializable {
899 private static final long serialVersionUID = 1L;
900 private final long currentTerm;
901 private final String votedFor;
903 public UpdateElectionTerm(long currentTerm, String votedFor) {
904 this.currentTerm = currentTerm;
905 this.votedFor = votedFor;
908 public long getCurrentTerm() {
912 public String getVotedFor() {
917 protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
919 public NonPersistentRaftDataProvider(){
924 * The way snapshotting works is,
926 * <li> RaftActor calls createSnapshot on the Shard
927 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
928 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
929 * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
930 * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
931 * in SaveSnapshotSuccess.
936 public void saveSnapshot(Object o) {
937 // Make saving Snapshot successful
943 void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
944 currentBehavior = behavior;
947 protected RaftActorBehavior getCurrentBehavior() {
948 return currentBehavior;