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 com.google.common.annotations.VisibleForTesting;
22 import com.google.common.base.Optional;
23 import com.google.common.base.Stopwatch;
24 import com.google.protobuf.ByteString;
25 import java.io.Serializable;
27 import org.opendaylight.controller.cluster.DataPersistenceProvider;
28 import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
29 import org.opendaylight.controller.cluster.notifications.RoleChanged;
30 import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
31 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
32 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
33 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
34 import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
35 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
36 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
37 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
38 import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
39 import org.opendaylight.controller.cluster.raft.behaviors.Follower;
40 import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
41 import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
42 import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
43 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
44 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
45 import org.opendaylight.controller.protobuff.messages.cluster.raft.AppendEntriesMessages;
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 LoggingAdapter LOG =
89 Logging.getLogger(getContext().system(), this);
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 volatile boolean hasSnapshotCaptureInitiated = false;
112 private Stopwatch recoveryTimer;
114 private int currentRecoveryBatchCount;
118 public RaftActor(String id, Map<String, String> peerAddresses) {
119 this(id, peerAddresses, Optional.<ConfigParams>absent());
122 public RaftActor(String id, Map<String, String> peerAddresses,
123 Optional<ConfigParams> configParams) {
125 context = new RaftActorContextImpl(this.getSelf(),
126 this.getContext(), id, new ElectionTermImpl(),
127 -1, -1, replicatedLog, peerAddresses,
128 (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()),
132 private void initRecoveryTimer() {
133 if(recoveryTimer == null) {
134 recoveryTimer = new Stopwatch();
135 recoveryTimer.start();
140 public void preStart() throws Exception {
141 LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
142 context.getConfigParams().getJournalRecoveryLogBatchSize());
148 public void handleRecover(Object message) {
149 if(persistence().isRecoveryApplicable()) {
150 if (message instanceof SnapshotOffer) {
151 onRecoveredSnapshot((SnapshotOffer) message);
152 } else if (message instanceof ReplicatedLogEntry) {
153 onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
154 } else if (message instanceof ApplyLogEntries) {
155 onRecoveredApplyLogEntries((ApplyLogEntries) message);
156 } else if (message instanceof DeleteEntries) {
157 replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
158 } else if (message instanceof UpdateElectionTerm) {
159 context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
160 ((UpdateElectionTerm) message).getVotedFor());
161 } else if (message instanceof RecoveryCompleted) {
162 onRecoveryCompletedMessage();
165 if (message instanceof RecoveryCompleted) {
166 // Delete all the messages from the akka journal so that we do not end up with consistency issues
167 // Note I am not using the dataPersistenceProvider and directly using the akka api here
168 deleteMessages(lastSequenceNr());
170 // Delete all the akka snapshots as they will not be needed
171 deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
173 onRecoveryComplete();
175 RaftActorBehavior oldBehavior = currentBehavior;
176 currentBehavior = new Follower(context);
177 handleBehaviorChange(oldBehavior, currentBehavior);
182 private void onRecoveredSnapshot(SnapshotOffer offer) {
183 if(LOG.isDebugEnabled()) {
184 LOG.debug("SnapshotOffer called..");
189 Snapshot snapshot = (Snapshot) offer.snapshot();
191 // Create a replicated log with the snapshot information
192 // The replicated log can be used later on to retrieve this snapshot
193 // when we need to install it on a peer
194 replicatedLog = new ReplicatedLogImpl(snapshot);
196 context.setReplicatedLog(replicatedLog);
197 context.setLastApplied(snapshot.getLastAppliedIndex());
198 context.setCommitIndex(snapshot.getLastAppliedIndex());
200 Stopwatch timer = new Stopwatch();
203 // Apply the snapshot to the actors state
204 applyRecoverySnapshot(ByteString.copyFrom(snapshot.getState()));
207 LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
208 replicatedLog.size(), persistenceId(), timer.toString(),
209 replicatedLog.snapshotIndex, replicatedLog.snapshotTerm);
212 private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
213 if(LOG.isDebugEnabled()) {
214 LOG.debug("Received ReplicatedLogEntry for recovery: {}", logEntry.getIndex());
217 replicatedLog.append(logEntry);
220 private void onRecoveredApplyLogEntries(ApplyLogEntries ale) {
221 if(LOG.isDebugEnabled()) {
222 LOG.debug("Received ApplyLogEntries for recovery, applying to state: {} to {}",
223 context.getLastApplied() + 1, ale.getToIndex());
226 for (long i = context.getLastApplied() + 1; i <= ale.getToIndex(); i++) {
227 batchRecoveredLogEntry(replicatedLog.get(i));
230 context.setLastApplied(ale.getToIndex());
231 context.setCommitIndex(ale.getToIndex());
234 private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
237 int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
238 if(currentRecoveryBatchCount == 0) {
239 startLogRecoveryBatch(batchSize);
242 appendRecoveredLogEntry(logEntry.getData());
244 if(++currentRecoveryBatchCount >= batchSize) {
245 endCurrentLogRecoveryBatch();
249 private void endCurrentLogRecoveryBatch() {
250 applyCurrentLogRecoveryBatch();
251 currentRecoveryBatchCount = 0;
254 private void onRecoveryCompletedMessage() {
255 if(currentRecoveryBatchCount > 0) {
256 endCurrentLogRecoveryBatch();
259 onRecoveryComplete();
261 String recoveryTime = "";
262 if(recoveryTimer != null) {
263 recoveryTimer.stop();
264 recoveryTime = " in " + recoveryTimer.toString();
265 recoveryTimer = null;
269 "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
270 "Persistence Id = " + persistenceId() +
271 " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
273 replicatedLog.lastIndex(), replicatedLog.snapshotIndex,
274 replicatedLog.snapshotTerm, replicatedLog.size());
276 RaftActorBehavior oldBehavior = currentBehavior;
277 currentBehavior = new Follower(context);
278 handleBehaviorChange(oldBehavior, currentBehavior);
281 @Override public void handleCommand(Object message) {
282 if (message instanceof ApplyState){
283 ApplyState applyState = (ApplyState) message;
285 if(LOG.isDebugEnabled()) {
286 LOG.debug("Applying state for log index {} data {}",
287 applyState.getReplicatedLogEntry().getIndex(),
288 applyState.getReplicatedLogEntry().getData());
291 applyState(applyState.getClientActor(), applyState.getIdentifier(),
292 applyState.getReplicatedLogEntry().getData());
294 } else if (message instanceof ApplyLogEntries){
295 ApplyLogEntries ale = (ApplyLogEntries) message;
296 if(LOG.isDebugEnabled()) {
297 LOG.debug("Persisting ApplyLogEntries with index={}", ale.getToIndex());
299 persistence().persist(new ApplyLogEntries(ale.getToIndex()), new Procedure<ApplyLogEntries>() {
301 public void apply(ApplyLogEntries param) throws Exception {
305 } else if(message instanceof ApplySnapshot ) {
306 Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
308 if(LOG.isDebugEnabled()) {
309 LOG.debug("ApplySnapshot called on Follower Actor " +
310 "snapshotIndex:{}, snapshotTerm:{}", snapshot.getLastAppliedIndex(),
311 snapshot.getLastAppliedTerm()
314 applySnapshot(ByteString.copyFrom(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");
331 long sequenceNumber = success.metadata().sequenceNr();
333 commitSnapshot(sequenceNumber);
335 } else if (message instanceof SaveSnapshotFailure) {
336 SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
338 LOG.info("saveSnapshotFailure.metadata():{}", saveSnapshotFailure.metadata().toString());
339 LOG.error(saveSnapshotFailure.cause(), "SaveSnapshotFailure received for snapshot Cause:");
341 context.getReplicatedLog().snapshotRollback();
343 LOG.info("Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
344 "snapshotIndex:{}, snapshotTerm:{}, log-size:{}",
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");
351 CaptureSnapshot cs = (CaptureSnapshot)message;
352 captureSnapshot = cs;
355 } else if (message instanceof CaptureSnapshotReply){
356 LOG.info("CaptureSnapshotReply received by actor");
357 CaptureSnapshotReply csr = (CaptureSnapshotReply) message;
359 ByteString stateInBytes = csr.getSnapshot();
360 LOG.info("CaptureSnapshotReply stateInBytes size:{}", stateInBytes.size());
361 handleCaptureSnapshotReply(stateInBytes);
364 if (!(message instanceof AppendEntriesMessages.AppendEntries)
365 && !(message instanceof AppendEntriesReply) && !(message instanceof SendHeartBeat)) {
366 if(LOG.isDebugEnabled()) {
367 LOG.debug("onReceiveCommand: message: {}", message.getClass());
371 RaftActorBehavior oldBehavior = currentBehavior;
372 currentBehavior = currentBehavior.handleMessage(getSender(), message);
374 handleBehaviorChange(oldBehavior, currentBehavior);
378 private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) {
379 if (oldBehavior != currentBehavior){
383 String oldBehaviorLeaderId = oldBehavior == null? null : oldBehavior.getLeaderId();
384 String oldBehaviorState = oldBehavior == null? null : oldBehavior.state().name();
386 // it can happen that the state has not changed but the leader has changed.
387 onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
389 if (getRoleChangeNotifier().isPresent() &&
390 (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
391 getRoleChangeNotifier().get().tell(
392 new RoleChanged(getId(), oldBehaviorState , currentBehavior.state().name()),
398 * When a derived RaftActor needs to persist something it must call
405 protected void persistData(final ActorRef clientActor, final String identifier,
406 final Payload data) {
408 ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
409 context.getReplicatedLog().lastIndex() + 1,
410 context.getTermInformation().getCurrentTerm(), data);
412 if(LOG.isDebugEnabled()) {
413 LOG.debug("Persist data {}", replicatedLogEntry);
416 final RaftActorContext raftContext = getRaftActorContext();
419 .appendAndPersist(replicatedLogEntry, new Procedure<ReplicatedLogEntry>() {
421 public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
423 // Increment the Commit Index and the Last Applied values
424 raftContext.setCommitIndex(replicatedLogEntry.getIndex());
425 raftContext.setLastApplied(replicatedLogEntry.getIndex());
427 // Apply the state immediately
428 applyState(clientActor, identifier, data);
430 // Send a ApplyLogEntries message so that we write the fact that we applied
431 // the state to durable storage
432 self().tell(new ApplyLogEntries((int) replicatedLogEntry.getIndex()), self());
434 // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot
435 if(!hasSnapshotCaptureInitiated){
436 raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(),
437 raftContext.getTermInformation().getCurrentTerm());
438 raftContext.getReplicatedLog().snapshotCommit();
440 LOG.debug("Skipping fake snapshotting for {} because real snapshotting is in progress", getId());
442 } else if (clientActor != null) {
443 // Send message for replication
444 currentBehavior.handleMessage(getSelf(),
445 new Replicate(clientActor, identifier,
453 protected String getId() {
454 return context.getId();
458 * Derived actors can call the isLeader method to check if the current
459 * RaftActor is the Leader or not
461 * @return true it this RaftActor is a Leader false otherwise
463 protected boolean isLeader() {
464 return context.getId().equals(currentBehavior.getLeaderId());
468 * Derived actor can call getLeader if they need a reference to the Leader.
469 * This would be useful for example in forwarding a request to an actor
470 * which is the leader
472 * @return A reference to the leader if known, null otherwise
474 protected ActorSelection getLeader(){
475 String leaderAddress = getLeaderAddress();
477 if(leaderAddress == null){
481 return context.actorSelection(leaderAddress);
486 * @return the current leader's id
488 protected String getLeaderId(){
489 return currentBehavior.getLeaderId();
492 protected RaftState getRaftState() {
493 return currentBehavior.state();
496 protected ReplicatedLogEntry getLastLogEntry() {
497 return replicatedLog.last();
500 protected Long getCurrentTerm(){
501 return context.getTermInformation().getCurrentTerm();
504 protected Long getCommitIndex(){
505 return context.getCommitIndex();
508 protected Long getLastApplied(){
509 return context.getLastApplied();
512 protected RaftActorContext getRaftActorContext() {
517 * setPeerAddress sets the address of a known peer at a later time.
519 * This is to account for situations where a we know that a peer
520 * exists but we do not know an address up-front. This may also be used in
521 * situations where a known peer starts off in a different location and we
522 * need to change it's address
524 * Note that if the peerId does not match the list of peers passed to
525 * this actor during construction an IllegalStateException will be thrown.
530 protected void setPeerAddress(String peerId, String peerAddress){
531 context.setPeerAddress(peerId, peerAddress);
534 protected void commitSnapshot(long sequenceNumber) {
535 context.getReplicatedLog().snapshotCommit();
537 // TODO: Not sure if we want to be this aggressive with trimming stuff
538 trimPersistentData(sequenceNumber);
542 * The applyState method will be called by the RaftActor when some data
543 * needs to be applied to the actor's state
545 * @param clientActor A reference to the client who sent this message. This
546 * is the same reference that was passed to persistData
547 * by the derived actor. clientActor may be null when
548 * the RaftActor is behaving as a follower or during
550 * @param identifier The identifier of the persisted data. This is also
551 * the same identifier that was passed to persistData by
552 * the derived actor. identifier may be null when
553 * the RaftActor is behaving as a follower or during
555 * @param data A piece of data that was persisted by the persistData call.
556 * This should NEVER be null.
558 protected abstract void applyState(ActorRef clientActor, String identifier,
562 * This method is called during recovery at the start of a batch of state entries. Derived
563 * classes should perform any initialization needed to start a batch.
565 protected abstract void startLogRecoveryBatch(int maxBatchSize);
568 * This method is called during recovery to append state data to the current batch. This method
569 * is called 1 or more times after {@link #startLogRecoveryBatch}.
571 * @param data the state data
573 protected abstract void appendRecoveredLogEntry(Payload data);
576 * This method is called during recovery to reconstruct the state of the actor.
578 * @param snapshot A snapshot of the state of the actor
580 protected abstract void applyRecoverySnapshot(ByteString snapshot);
583 * This method is called during recovery at the end of a batch to apply the current batched
584 * log entries. This method is called after {@link #appendRecoveredLogEntry}.
586 protected abstract void applyCurrentLogRecoveryBatch();
589 * This method is called when recovery is complete.
591 protected abstract void onRecoveryComplete();
594 * This method will be called by the RaftActor when a snapshot needs to be
595 * created. The derived actor should respond with its current state.
597 * During recovery the state that is returned by the derived actor will
598 * be passed back to it by calling the applySnapshot method
600 * @return The current state of the actor
602 protected abstract void createSnapshot();
605 * This method can be called at any other point during normal
606 * operations when the derived actor is out of sync with it's peers
607 * and the only way to bring it in sync is by applying a snapshot
609 * @param snapshot A snapshot of the state of the actor
611 protected abstract void applySnapshot(ByteString snapshot);
614 * This method will be called by the RaftActor when the state of the
615 * RaftActor changes. The derived actor can then use methods like
616 * isLeader or getLeader to do something useful
618 protected abstract void onStateChanged();
620 protected abstract DataPersistenceProvider persistence();
623 * Notifier Actor for this RaftActor to notify when a role change happens
624 * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
626 protected abstract Optional<ActorRef> getRoleChangeNotifier();
628 protected void onLeaderChanged(String oldLeader, String newLeader){};
630 private void trimPersistentData(long sequenceNumber) {
631 // Trim akka snapshots
632 // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
633 // For now guessing that it is ANDed.
634 persistence().deleteSnapshots(new SnapshotSelectionCriteria(
635 sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
638 persistence().deleteMessages(sequenceNumber);
641 private String getLeaderAddress(){
643 return getSelf().path().toString();
645 String leaderId = currentBehavior.getLeaderId();
646 if (leaderId == null) {
649 String peerAddress = context.getPeerAddress(leaderId);
650 if(LOG.isDebugEnabled()) {
651 LOG.debug("getLeaderAddress leaderId = {} peerAddress = {}",
652 leaderId, peerAddress);
658 private void handleCaptureSnapshotReply(ByteString stateInBytes) {
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(stateInBytes.toByteArray(),
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:{}", 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:{}", 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(stateInBytes));
687 captureSnapshot = null;
688 hasSnapshotCaptureInitiated = false;
691 protected boolean hasFollowers(){
692 return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
695 private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
697 private static final int DATA_SIZE_DIVIDER = 5;
698 private long dataSizeSinceLastSnapshot = 0;
700 public ReplicatedLogImpl(Snapshot snapshot) {
701 super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
702 snapshot.getUnAppliedEntries());
705 public ReplicatedLogImpl() {
709 @Override public void removeFromAndPersist(long logEntryIndex) {
710 int adjustedIndex = adjustedIndex(logEntryIndex);
712 if (adjustedIndex < 0) {
716 // FIXME: Maybe this should be done after the command is saved
717 journal.subList(adjustedIndex , journal.size()).clear();
719 persistence().persist(new DeleteEntries(adjustedIndex), new Procedure<DeleteEntries>(){
721 @Override public void apply(DeleteEntries param)
723 //FIXME : Doing nothing for now
725 for(ReplicatedLogEntry entry : journal){
726 dataSize += entry.size();
732 @Override public void appendAndPersist(
733 final ReplicatedLogEntry replicatedLogEntry) {
734 appendAndPersist(replicatedLogEntry, null);
738 public int dataSize() {
742 public void appendAndPersist(
743 final ReplicatedLogEntry replicatedLogEntry,
744 final Procedure<ReplicatedLogEntry> callback) {
746 if(LOG.isDebugEnabled()) {
747 LOG.debug("Append log entry and persist {} ", replicatedLogEntry);
750 // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
751 journal.add(replicatedLogEntry);
753 // When persisting events with persist it is guaranteed that the
754 // persistent actor will not receive further commands between the
755 // persist call and the execution(s) of the associated event
756 // handler. This also holds for multiple persist calls in context
757 // of a single command.
758 persistence().persist(replicatedLogEntry,
759 new Procedure<ReplicatedLogEntry>() {
761 public void apply(ReplicatedLogEntry evt) throws Exception {
762 int logEntrySize = replicatedLogEntry.size();
764 dataSize += logEntrySize;
765 long dataSizeForCheck = dataSize;
767 dataSizeSinceLastSnapshot += logEntrySize;
768 long journalSize = lastIndex()+1;
770 if(!hasFollowers()) {
771 // When we do not have followers we do not maintain an in-memory log
772 // due to this the journalSize will never become anything close to the
773 // snapshot batch count. In fact will mostly be 1.
774 // Similarly since the journal's dataSize depends on the entries in the
775 // journal the journal's dataSize will never reach a value close to the
777 // By maintaining the dataSize outside the journal we are tracking essentially
778 // what we have written to the disk however since we no longer are in
779 // need of doing a snapshot just for the sake of freeing up memory we adjust
780 // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
781 // as if we were maintaining a real snapshot
782 dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
785 long dataThreshold = Runtime.getRuntime().totalMemory() *
786 getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
788 // when a snaphsot is being taken, captureSnapshot != null
789 if (hasSnapshotCaptureInitiated == false &&
790 ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
791 dataSizeForCheck > dataThreshold)) {
793 dataSizeSinceLastSnapshot = 0;
795 LOG.info("Initiating Snapshot Capture..");
796 long lastAppliedIndex = -1;
797 long lastAppliedTerm = -1;
799 ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
800 if (!hasFollowers()) {
801 lastAppliedIndex = replicatedLogEntry.getIndex();
802 lastAppliedTerm = replicatedLogEntry.getTerm();
803 } else if (lastAppliedEntry != null) {
804 lastAppliedIndex = lastAppliedEntry.getIndex();
805 lastAppliedTerm = lastAppliedEntry.getTerm();
808 if(LOG.isDebugEnabled()) {
809 LOG.debug("Snapshot Capture logSize: {}", journal.size());
810 LOG.debug("Snapshot Capture lastApplied:{} ",
811 context.getLastApplied());
812 LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
813 LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
816 // send a CaptureSnapshot to self to make the expensive operation async.
817 getSelf().tell(new CaptureSnapshot(
818 lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm),
820 hasSnapshotCaptureInitiated = true;
822 if(callback != null){
823 callback.apply(replicatedLogEntry);
832 static class DeleteEntries implements Serializable {
833 private static final long serialVersionUID = 1L;
834 private final int fromIndex;
836 public DeleteEntries(int fromIndex) {
837 this.fromIndex = fromIndex;
840 public int getFromIndex() {
846 private class ElectionTermImpl implements ElectionTerm {
848 * Identifier of the actor whose election term information this is
850 private long currentTerm = 0;
851 private String votedFor = null;
854 public long getCurrentTerm() {
859 public String getVotedFor() {
863 @Override public void update(long currentTerm, String votedFor) {
864 if(LOG.isDebugEnabled()) {
865 LOG.debug("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
867 this.currentTerm = currentTerm;
868 this.votedFor = votedFor;
872 public void updateAndPersist(long currentTerm, String votedFor){
873 update(currentTerm, votedFor);
874 // FIXME : Maybe first persist then update the state
875 persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure<UpdateElectionTerm>(){
877 @Override public void apply(UpdateElectionTerm param)
885 static class UpdateElectionTerm implements Serializable {
886 private static final long serialVersionUID = 1L;
887 private final long currentTerm;
888 private final String votedFor;
890 public UpdateElectionTerm(long currentTerm, String votedFor) {
891 this.currentTerm = currentTerm;
892 this.votedFor = votedFor;
895 public long getCurrentTerm() {
899 public String getVotedFor() {
904 protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
906 public NonPersistentRaftDataProvider(){
911 * The way snapshotting works is,
913 * <li> RaftActor calls createSnapshot on the Shard
914 * <li> Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
915 * <li> When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
916 * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
917 * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
918 * in SaveSnapshotSuccess.
923 public void saveSnapshot(Object o) {
924 // Make saving Snapshot successful
930 void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
931 currentBehavior = behavior;
934 protected RaftActorBehavior getCurrentBehavior() {
935 return currentBehavior;