X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?a=blobdiff_plain;ds=inline;f=opendaylight%2Fmd-sal%2Fsal-akka-raft%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fraft%2FRaftActor.java;h=2459c2ff8b1764d3cd3b56be90fc7ea5191d65b6;hb=2a89ae48921724ef5a4ab42dcff6afc74c5b0a4a;hp=988789b4011e4f3ba9e9e9abe6b0c3009704c810;hpb=bff72932e293fe108bc749bd168dbc415088584e;p=controller.git
diff --git a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java
index 988789b401..3ec8cc5c58 100644
--- a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java
+++ b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActor.java
@@ -10,32 +10,37 @@ package org.opendaylight.controller.cluster.raft;
import akka.actor.ActorRef;
import akka.actor.ActorSelection;
-import akka.event.Logging;
-import akka.event.LoggingAdapter;
import akka.japi.Procedure;
import akka.persistence.RecoveryCompleted;
import akka.persistence.SaveSnapshotFailure;
import akka.persistence.SaveSnapshotSuccess;
import akka.persistence.SnapshotOffer;
import akka.persistence.SnapshotSelectionCriteria;
-import akka.persistence.UntypedPersistentActor;
+import com.google.common.annotations.VisibleForTesting;
+import com.google.common.base.Optional;
+import com.google.common.base.Stopwatch;
+import com.google.protobuf.ByteString;
+import java.io.Serializable;
+import java.util.Map;
+import java.util.concurrent.TimeUnit;
+import org.opendaylight.controller.cluster.DataPersistenceProvider;
+import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor;
+import org.opendaylight.controller.cluster.notifications.RoleChanged;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
+import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
+import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshotReply;
import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
-import com.google.common.base.Optional;
-import org.opendaylight.controller.cluster.raft.behaviors.Candidate;
+import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
+import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior;
import org.opendaylight.controller.cluster.raft.behaviors.Follower;
-import org.opendaylight.controller.cluster.raft.behaviors.Leader;
import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior;
-import org.opendaylight.controller.cluster.raft.client.messages.AddRaftPeer;
import org.opendaylight.controller.cluster.raft.client.messages.FindLeader;
import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply;
-import org.opendaylight.controller.cluster.raft.client.messages.RemoveRaftPeer;
import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
-
-import java.io.Serializable;
-import java.util.List;
-import java.util.Map;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
/**
* RaftActor encapsulates a state machine that needs to be kept synchronized
@@ -77,9 +82,11 @@ import java.util.Map;
*
when a snapshot should be saved
*
*/
-public abstract class RaftActor extends UntypedPersistentActor {
- protected final LoggingAdapter LOG =
- Logging.getLogger(getContext().system(), this);
+public abstract class RaftActor extends AbstractUntypedPersistentActor {
+
+ private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis
+
+ protected final Logger LOG = LoggerFactory.getLogger(getClass());
/**
* The current state determines the current behavior of a RaftActor
@@ -91,13 +98,18 @@ public abstract class RaftActor extends UntypedPersistentActor {
* This context should NOT be passed directly to any other actor it is
* only to be consumed by the RaftActorBehaviors
*/
- private RaftActorContext context;
+ private final RaftActorContext context;
/**
* The in-memory journal
*/
private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl();
+ private CaptureSnapshot captureSnapshot = null;
+
+ private Stopwatch recoveryTimer;
+
+ private int currentRecoveryBatchCount;
public RaftActor(String id, Map peerAddresses) {
this(id, peerAddresses, Optional.absent());
@@ -113,58 +125,205 @@ public abstract class RaftActor extends UntypedPersistentActor {
LOG);
}
- @Override public void onReceiveRecover(Object message) {
- if (message instanceof SnapshotOffer) {
- LOG.debug("SnapshotOffer called..");
- SnapshotOffer offer = (SnapshotOffer) message;
- Snapshot snapshot = (Snapshot) offer.snapshot();
+ private void initRecoveryTimer() {
+ if(recoveryTimer == null) {
+ recoveryTimer = Stopwatch.createStarted();
+ }
+ }
+
+ @Override
+ public void preStart() throws Exception {
+ LOG.info("Starting recovery for {} with journal batch size {}", persistenceId(),
+ context.getConfigParams().getJournalRecoveryLogBatchSize());
- // Create a replicated log with the snapshot information
- // The replicated log can be used later on to retrieve this snapshot
- // when we need to install it on a peer
- replicatedLog = new ReplicatedLogImpl(snapshot);
+ super.preStart();
+ }
- context.setReplicatedLog(replicatedLog);
+ @Override
+ public void handleRecover(Object message) {
+ if(persistence().isRecoveryApplicable()) {
+ if (message instanceof SnapshotOffer) {
+ onRecoveredSnapshot((SnapshotOffer) message);
+ } else if (message instanceof ReplicatedLogEntry) {
+ onRecoveredJournalLogEntry((ReplicatedLogEntry) message);
+ } else if (message instanceof ApplyLogEntries) {
+ onRecoveredApplyLogEntries((ApplyLogEntries) message);
+ } else if (message instanceof DeleteEntries) {
+ replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
+ } else if (message instanceof UpdateElectionTerm) {
+ context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(),
+ ((UpdateElectionTerm) message).getVotedFor());
+ } else if (message instanceof RecoveryCompleted) {
+ onRecoveryCompletedMessage();
+ }
+ } else {
+ if (message instanceof RecoveryCompleted) {
+ // Delete all the messages from the akka journal so that we do not end up with consistency issues
+ // Note I am not using the dataPersistenceProvider and directly using the akka api here
+ deleteMessages(lastSequenceNr());
- LOG.debug("Applied snapshot to replicatedLog. " +
- "snapshotIndex={}, snapshotTerm={}, journal-size={}",
- replicatedLog.snapshotIndex, replicatedLog.snapshotTerm,
- replicatedLog.size());
+ // Delete all the akka snapshots as they will not be needed
+ deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue()));
- // Apply the snapshot to the actors state
- applySnapshot(snapshot.getState());
+ onRecoveryComplete();
- } else if (message instanceof ReplicatedLogEntry) {
- replicatedLog.append((ReplicatedLogEntry) message);
- } else if (message instanceof DeleteEntries) {
- replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex());
- } else if (message instanceof UpdateElectionTerm) {
- context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor());
- } else if (message instanceof RecoveryCompleted) {
- LOG.debug(
- "RecoveryCompleted - Switching actor to Follower - " +
- "Last index in log:{}, snapshotIndex={}, snapshotTerm={}, " +
- "journal-size={}",
- replicatedLog.lastIndex(), replicatedLog.snapshotIndex,
- replicatedLog.snapshotTerm, replicatedLog.size());
- currentBehavior = switchBehavior(RaftState.Follower);
- onStateChanged();
+ initializeBehavior();
+ }
+ }
+ }
+
+ private void onRecoveredSnapshot(SnapshotOffer offer) {
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: SnapshotOffer called..", persistenceId());
+ }
+
+ initRecoveryTimer();
+
+ Snapshot snapshot = (Snapshot) offer.snapshot();
+
+ // Create a replicated log with the snapshot information
+ // The replicated log can be used later on to retrieve this snapshot
+ // when we need to install it on a peer
+ replicatedLog = new ReplicatedLogImpl(snapshot);
+
+ context.setReplicatedLog(replicatedLog);
+ context.setLastApplied(snapshot.getLastAppliedIndex());
+ context.setCommitIndex(snapshot.getLastAppliedIndex());
+
+ Stopwatch timer = Stopwatch.createStarted();
+
+ // Apply the snapshot to the actors state
+ applyRecoverySnapshot(snapshot.getState());
+
+ timer.stop();
+ LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" +
+ replicatedLog.size(), persistenceId(), timer.toString(),
+ replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm());
+ }
+
+ private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) {
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex());
+ }
+
+ replicatedLog.append(logEntry);
+ }
+
+ private void onRecoveredApplyLogEntries(ApplyLogEntries ale) {
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}",
+ persistenceId(), context.getLastApplied() + 1, ale.getToIndex());
+ }
+
+ for (long i = context.getLastApplied() + 1; i <= ale.getToIndex(); i++) {
+ batchRecoveredLogEntry(replicatedLog.get(i));
+ }
+
+ context.setLastApplied(ale.getToIndex());
+ context.setCommitIndex(ale.getToIndex());
+ }
+
+ private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) {
+ initRecoveryTimer();
+
+ int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize();
+ if(currentRecoveryBatchCount == 0) {
+ startLogRecoveryBatch(batchSize);
+ }
+
+ appendRecoveredLogEntry(logEntry.getData());
+
+ if(++currentRecoveryBatchCount >= batchSize) {
+ endCurrentLogRecoveryBatch();
+ }
+ }
+
+ private void endCurrentLogRecoveryBatch() {
+ applyCurrentLogRecoveryBatch();
+ currentRecoveryBatchCount = 0;
+ }
+
+ private void onRecoveryCompletedMessage() {
+ if(currentRecoveryBatchCount > 0) {
+ endCurrentLogRecoveryBatch();
}
+
+ onRecoveryComplete();
+
+ String recoveryTime = "";
+ if(recoveryTimer != null) {
+ recoveryTimer.stop();
+ recoveryTime = " in " + recoveryTimer.toString();
+ recoveryTimer = null;
+ }
+
+ LOG.info(
+ "Recovery completed" + recoveryTime + " - Switching actor to Follower - " +
+ "Persistence Id = " + persistenceId() +
+ " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " +
+ "journal-size={}",
+ replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(),
+ replicatedLog.getSnapshotTerm(), replicatedLog.size());
+
+ initializeBehavior();
}
- @Override public void onReceiveCommand(Object message) {
+ protected void initializeBehavior(){
+ changeCurrentBehavior(new Follower(context));
+ }
+
+ protected void changeCurrentBehavior(RaftActorBehavior newBehavior){
+ RaftActorBehavior oldBehavior = currentBehavior;
+ currentBehavior = newBehavior;
+ handleBehaviorChange(oldBehavior, currentBehavior);
+ }
+
+ @Override public void handleCommand(Object message) {
if (message instanceof ApplyState){
ApplyState applyState = (ApplyState) message;
- LOG.debug("Applying state for log index {} data {}",
- applyState.getReplicatedLogEntry().getIndex(),
- applyState.getReplicatedLogEntry().getData());
+ long elapsedTime = (System.nanoTime() - applyState.getStartTime());
+ if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
+ LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
+ TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
+ }
+
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Applying state for log index {} data {}",
+ persistenceId(), applyState.getReplicatedLogEntry().getIndex(),
+ applyState.getReplicatedLogEntry().getData());
+ }
applyState(applyState.getClientActor(), applyState.getIdentifier(),
applyState.getReplicatedLogEntry().getData());
+ } else if (message instanceof ApplyLogEntries){
+ ApplyLogEntries ale = (ApplyLogEntries) message;
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), ale.getToIndex());
+ }
+ persistence().persist(new ApplyLogEntries(ale.getToIndex()), new Procedure() {
+ @Override
+ public void apply(ApplyLogEntries param) throws Exception {
+ }
+ });
+
} else if(message instanceof ApplySnapshot ) {
- applySnapshot(((ApplySnapshot) message).getSnapshot());
+ Snapshot snapshot = ((ApplySnapshot) message).getSnapshot();
+
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: ApplySnapshot called on Follower Actor " +
+ "snapshotIndex:{}, snapshotTerm:{}", persistenceId(), snapshot.getLastAppliedIndex(),
+ snapshot.getLastAppliedTerm()
+ );
+ }
+
+ applySnapshot(snapshot.getState());
+
+ //clears the followers log, sets the snapshot index to ensure adjusted-index works
+ replicatedLog = new ReplicatedLogImpl(snapshot);
+ context.setReplicatedLog(replicatedLog);
+ context.setLastApplied(snapshot.getLastAppliedIndex());
} else if (message instanceof FindLeader) {
getSender().tell(
@@ -174,50 +333,63 @@ public abstract class RaftActor extends UntypedPersistentActor {
} else if (message instanceof SaveSnapshotSuccess) {
SaveSnapshotSuccess success = (SaveSnapshotSuccess) message;
+ LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId());
- // TODO: Not sure if we want to be this aggressive with trimming stuff
- trimPersistentData(success.metadata().sequenceNr());
+ long sequenceNumber = success.metadata().sequenceNr();
+
+ commitSnapshot(sequenceNumber);
} else if (message instanceof SaveSnapshotFailure) {
+ SaveSnapshotFailure saveSnapshotFailure = (SaveSnapshotFailure) message;
+
+ LOG.error("{}: SaveSnapshotFailure received for snapshot Cause:",
+ persistenceId(), saveSnapshotFailure.cause());
- // TODO: Handle failure in saving the snapshot
+ context.getReplicatedLog().snapshotRollback();
- } else if (message instanceof AddRaftPeer){
+ LOG.info("{}: Replicated Log rollbacked. Snapshot will be attempted in the next cycle." +
+ "snapshotIndex:{}, snapshotTerm:{}, log-size:{}", persistenceId(),
+ context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().getSnapshotTerm(),
+ context.getReplicatedLog().size());
- // FIXME : Do not add raft peers like this.
- // When adding a new Peer we have to ensure that the a majority of
- // the peers know about the new Peer. Doing it this way may cause
- // a situation where multiple Leaders may emerge
- AddRaftPeer arp = (AddRaftPeer)message;
- context.addToPeers(arp.getName(), arp.getAddress());
+ } else if (message instanceof CaptureSnapshot) {
+ LOG.info("{}: CaptureSnapshot received by actor", persistenceId());
- } else if (message instanceof RemoveRaftPeer){
+ if(captureSnapshot == null) {
+ captureSnapshot = (CaptureSnapshot)message;
+ createSnapshot();
+ }
- RemoveRaftPeer rrp = (RemoveRaftPeer)message;
- context.removePeer(rrp.getName());
+ } else if (message instanceof CaptureSnapshotReply){
+ handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot());
} else {
-
- RaftState state =
- currentBehavior.handleMessage(getSender(), message);
RaftActorBehavior oldBehavior = currentBehavior;
- currentBehavior = switchBehavior(state);
- if(oldBehavior != currentBehavior){
- onStateChanged();
- }
+ currentBehavior = currentBehavior.handleMessage(getSender(), message);
+
+ handleBehaviorChange(oldBehavior, currentBehavior);
}
}
- public java.util.Set getPeers() {
- return context.getPeerAddresses().keySet();
- }
+ private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) {
+ if (oldBehavior != currentBehavior){
+ onStateChanged();
+ }
- protected String getReplicatedLogState() {
- return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex()
- + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm()
- + ", im-mem journal size=" + context.getReplicatedLog().size();
- }
+ String oldBehaviorLeaderId = oldBehavior == null? null : oldBehavior.getLeaderId();
+ String oldBehaviorState = oldBehavior == null? null : oldBehavior.state().name();
+ // it can happen that the state has not changed but the leader has changed.
+ onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId());
+
+ if (getRoleChangeNotifier().isPresent() &&
+ (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) {
+ getRoleChangeNotifier().get().tell(
+ new RoleChanged(getId(), oldBehaviorState , currentBehavior.state().name()),
+ getSelf());
+ }
+ }
/**
* When a derived RaftActor needs to persist something it must call
@@ -227,18 +399,54 @@ public abstract class RaftActor extends UntypedPersistentActor {
* @param identifier
* @param data
*/
- protected void persistData(ActorRef clientActor, String identifier,
- Payload data) {
+ protected void persistData(final ActorRef clientActor, final String identifier,
+ final Payload data) {
ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry(
context.getReplicatedLog().lastIndex() + 1,
context.getTermInformation().getCurrentTerm(), data);
- LOG.debug("Persist data {}", replicatedLogEntry);
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Persist data {}", persistenceId(), replicatedLogEntry);
+ }
+
+ final RaftActorContext raftContext = getRaftActorContext();
replicatedLog
- .appendAndPersist(clientActor, identifier, replicatedLogEntry);
- }
+ .appendAndPersist(replicatedLogEntry, new Procedure() {
+ @Override
+ public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception {
+ if(!hasFollowers()){
+ // Increment the Commit Index and the Last Applied values
+ raftContext.setCommitIndex(replicatedLogEntry.getIndex());
+ raftContext.setLastApplied(replicatedLogEntry.getIndex());
+
+ // Apply the state immediately
+ applyState(clientActor, identifier, data);
+
+ // Send a ApplyLogEntries message so that we write the fact that we applied
+ // the state to durable storage
+ self().tell(new ApplyLogEntries((int) replicatedLogEntry.getIndex()), self());
+
+ // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot
+ if(!context.isSnapshotCaptureInitiated()){
+ raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(),
+ raftContext.getTermInformation().getCurrentTerm());
+ raftContext.getReplicatedLog().snapshotCommit();
+ } else {
+ LOG.debug("{}: Skipping fake snapshotting for {} because real snapshotting is in progress",
+ persistenceId(), getId());
+ }
+ } else if (clientActor != null) {
+ // Send message for replication
+ currentBehavior.handleMessage(getSelf(),
+ new Replicate(clientActor, identifier,
+ replicatedLogEntry)
+ );
+ }
+
+ }
+ }); }
protected String getId() {
return context.getId();
@@ -264,6 +472,10 @@ public abstract class RaftActor extends UntypedPersistentActor {
protected ActorSelection getLeader(){
String leaderAddress = getLeaderAddress();
+ if(leaderAddress == null){
+ return null;
+ }
+
return context.actorSelection(leaderAddress);
}
@@ -295,6 +507,10 @@ public abstract class RaftActor extends UntypedPersistentActor {
return context.getLastApplied();
}
+ protected RaftActorContext getRaftActorContext() {
+ return context;
+ }
+
/**
* setPeerAddress sets the address of a known peer at a later time.
*
@@ -313,7 +529,12 @@ public abstract class RaftActor extends UntypedPersistentActor {
context.setPeerAddress(peerId, peerAddress);
}
+ protected void commitSnapshot(long sequenceNumber) {
+ context.getReplicatedLog().snapshotCommit();
+ // TODO: Not sure if we want to be this aggressive with trimming stuff
+ trimPersistentData(sequenceNumber);
+ }
/**
* The applyState method will be called by the RaftActor when some data
@@ -335,6 +556,38 @@ public abstract class RaftActor extends UntypedPersistentActor {
protected abstract void applyState(ActorRef clientActor, String identifier,
Object data);
+ /**
+ * This method is called during recovery at the start of a batch of state entries. Derived
+ * classes should perform any initialization needed to start a batch.
+ */
+ protected abstract void startLogRecoveryBatch(int maxBatchSize);
+
+ /**
+ * This method is called during recovery to append state data to the current batch. This method
+ * is called 1 or more times after {@link #startLogRecoveryBatch}.
+ *
+ * @param data the state data
+ */
+ protected abstract void appendRecoveredLogEntry(Payload data);
+
+ /**
+ * This method is called during recovery to reconstruct the state of the actor.
+ *
+ * @param snapshotBytes A snapshot of the state of the actor
+ */
+ protected abstract void applyRecoverySnapshot(byte[] snapshotBytes);
+
+ /**
+ * This method is called during recovery at the end of a batch to apply the current batched
+ * log entries. This method is called after {@link #appendRecoveredLogEntry}.
+ */
+ protected abstract void applyCurrentLogRecoveryBatch();
+
+ /**
+ * This method is called when recovery is complete.
+ */
+ protected abstract void onRecoveryComplete();
+
/**
* This method will be called by the RaftActor when a snapshot needs to be
* created. The derived actor should respond with its current state.
@@ -344,19 +597,16 @@ public abstract class RaftActor extends UntypedPersistentActor {
*
* @return The current state of the actor
*/
- protected abstract Object createSnapshot();
+ protected abstract void createSnapshot();
/**
- * This method will be called by the RaftActor during recovery to
- * reconstruct the state of the actor.
- *
- * This method may also be called at any other point during normal
+ * This method can be called at any other point during normal
* operations when the derived actor is out of sync with it's peers
* and the only way to bring it in sync is by applying a snapshot
*
- * @param snapshot A snapshot of the state of the actor
+ * @param snapshotBytes A snapshot of the state of the actor
*/
- protected abstract void applySnapshot(Object snapshot);
+ protected abstract void applySnapshot(byte[] snapshotBytes);
/**
* This method will be called by the RaftActor when the state of the
@@ -365,47 +615,25 @@ public abstract class RaftActor extends UntypedPersistentActor {
*/
protected abstract void onStateChanged();
- private RaftActorBehavior switchBehavior(RaftState state) {
- if (currentBehavior != null) {
- if (currentBehavior.state() == state) {
- return currentBehavior;
- }
- LOG.info("Switching from state " + currentBehavior.state() + " to "
- + state);
-
- try {
- currentBehavior.close();
- } catch (Exception e) {
- LOG.error(e,
- "Failed to close behavior : " + currentBehavior.state());
- }
-
- } else {
- LOG.info("Switching behavior to " + state);
- }
- RaftActorBehavior behavior = null;
- if (state == RaftState.Candidate) {
- behavior = new Candidate(context);
- } else if (state == RaftState.Follower) {
- behavior = new Follower(context);
- } else {
- behavior = new Leader(context);
- }
-
+ protected abstract DataPersistenceProvider persistence();
+ /**
+ * Notifier Actor for this RaftActor to notify when a role change happens
+ * @return ActorRef - ActorRef of the notifier or Optional.absent if none.
+ */
+ protected abstract Optional getRoleChangeNotifier();
- return behavior;
- }
+ protected void onLeaderChanged(String oldLeader, String newLeader){};
private void trimPersistentData(long sequenceNumber) {
// Trim akka snapshots
// FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied
// For now guessing that it is ANDed.
- deleteSnapshots(new SnapshotSelectionCriteria(
+ persistence().deleteSnapshots(new SnapshotSelectionCriteria(
sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000));
// Trim akka journal
- deleteMessages(sequenceNumber);
+ persistence().deleteMessages(sequenceNumber);
}
private String getLeaderAddress(){
@@ -417,18 +645,70 @@ public abstract class RaftActor extends UntypedPersistentActor {
return null;
}
String peerAddress = context.getPeerAddress(leaderId);
- LOG.debug("getLeaderAddress leaderId = " + leaderId + " peerAddress = "
- + peerAddress);
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: getLeaderAddress leaderId = {} peerAddress = {}",
+ persistenceId(), leaderId, peerAddress);
+ }
return peerAddress;
}
+ private void handleCaptureSnapshotReply(byte[] snapshotBytes) {
+ LOG.info("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length);
+
+ // create a snapshot object from the state provided and save it
+ // when snapshot is saved async, SaveSnapshotSuccess is raised.
+
+ Snapshot sn = Snapshot.create(snapshotBytes,
+ context.getReplicatedLog().getFrom(captureSnapshot.getLastAppliedIndex() + 1),
+ captureSnapshot.getLastIndex(), captureSnapshot.getLastTerm(),
+ captureSnapshot.getLastAppliedIndex(), captureSnapshot.getLastAppliedTerm());
+
+ persistence().saveSnapshot(sn);
+
+ LOG.info("{}: Persisting of snapshot done:{}", persistenceId(), sn.getLogMessage());
+
+ long dataThreshold = Runtime.getRuntime().totalMemory() *
+ getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
+ if (context.getReplicatedLog().dataSize() > dataThreshold) {
+ // if memory is less, clear the log based on lastApplied.
+ // this could/should only happen if one of the followers is down
+ // as normally we keep removing from the log when its replicated to all.
+ context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getLastAppliedIndex(),
+ captureSnapshot.getLastAppliedTerm());
+
+ } else {
+ // clear the log based on replicatedToAllIndex
+ context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getReplicatedToAllIndex(),
+ captureSnapshot.getReplicatedToAllTerm());
+ }
+ getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex());
+
+ LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex:{} " +
+ "and term:{}", persistenceId(), captureSnapshot.getLastAppliedIndex(),
+ captureSnapshot.getLastAppliedTerm());
+
+ if (isLeader() && captureSnapshot.isInstallSnapshotInitiated()) {
+ // this would be call straight to the leader and won't initiate in serialization
+ currentBehavior.handleMessage(getSelf(), new SendInstallSnapshot(
+ ByteString.copyFrom(snapshotBytes)));
+ }
+
+ captureSnapshot = null;
+ context.setSnapshotCaptureInitiated(false);
+ }
+
+ protected boolean hasFollowers(){
+ return getRaftActorContext().getPeerAddresses().keySet().size() > 0;
+ }
private class ReplicatedLogImpl extends AbstractReplicatedLogImpl {
+ private static final int DATA_SIZE_DIVIDER = 5;
+ private long dataSizeSinceLastSnapshot = 0;
+
public ReplicatedLogImpl(Snapshot snapshot) {
- super(snapshot.getState(),
- snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
+ super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(),
snapshot.getUnAppliedEntries());
}
@@ -446,25 +726,33 @@ public abstract class RaftActor extends UntypedPersistentActor {
// FIXME: Maybe this should be done after the command is saved
journal.subList(adjustedIndex , journal.size()).clear();
- persist(new DeleteEntries(adjustedIndex), new Procedure(){
+ persistence().persist(new DeleteEntries(adjustedIndex), new Procedure() {
- @Override public void apply(DeleteEntries param)
- throws Exception {
+ @Override
+ public void apply(DeleteEntries param)
+ throws Exception {
//FIXME : Doing nothing for now
+ dataSize = 0;
+ for (ReplicatedLogEntry entry : journal) {
+ dataSize += entry.size();
+ }
}
});
}
@Override public void appendAndPersist(
final ReplicatedLogEntry replicatedLogEntry) {
- appendAndPersist(null, null, replicatedLogEntry);
+ appendAndPersist(replicatedLogEntry, null);
}
- public void appendAndPersist(final ActorRef clientActor,
- final String identifier,
- final ReplicatedLogEntry replicatedLogEntry) {
- context.getLogger().debug(
- "Append log entry and persist {} ", replicatedLogEntry);
+ public void appendAndPersist(
+ final ReplicatedLogEntry replicatedLogEntry,
+ final Procedure callback) {
+
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Append log entry and persist {} ", persistenceId(), replicatedLogEntry);
+ }
+
// FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs
journal.add(replicatedLogEntry);
@@ -473,53 +761,79 @@ public abstract class RaftActor extends UntypedPersistentActor {
// persist call and the execution(s) of the associated event
// handler. This also holds for multiple persist calls in context
// of a single command.
- persist(replicatedLogEntry,
+ persistence().persist(replicatedLogEntry,
new Procedure() {
+ @Override
public void apply(ReplicatedLogEntry evt) throws Exception {
- // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned.
- if (journal.size() > context.getConfigParams().getSnapshotBatchCount()) {
- LOG.info("Initiating Snapshot Capture..");
+ int logEntrySize = replicatedLogEntry.size();
+
+ dataSize += logEntrySize;
+ long dataSizeForCheck = dataSize;
+
+ dataSizeSinceLastSnapshot += logEntrySize;
+ long journalSize = lastIndex() + 1;
+
+ if(!hasFollowers()) {
+ // When we do not have followers we do not maintain an in-memory log
+ // due to this the journalSize will never become anything close to the
+ // snapshot batch count. In fact will mostly be 1.
+ // Similarly since the journal's dataSize depends on the entries in the
+ // journal the journal's dataSize will never reach a value close to the
+ // memory threshold.
+ // By maintaining the dataSize outside the journal we are tracking essentially
+ // what we have written to the disk however since we no longer are in
+ // need of doing a snapshot just for the sake of freeing up memory we adjust
+ // the real size of data by the DATA_SIZE_DIVIDER so that we do not snapshot as often
+ // as if we were maintaining a real snapshot
+ dataSizeForCheck = dataSizeSinceLastSnapshot / DATA_SIZE_DIVIDER;
+ }
+
+ long dataThreshold = Runtime.getRuntime().totalMemory() *
+ getRaftActorContext().getConfigParams().getSnapshotDataThresholdPercentage() / 100;
+
+ // when a snaphsot is being taken, captureSnapshot != null
+ if (!context.isSnapshotCaptureInitiated() &&
+ ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 ||
+ dataSizeForCheck > dataThreshold)) {
+
+ dataSizeSinceLastSnapshot = 0;
+
+ LOG.info("{}: Initiating Snapshot Capture, journalSize = {}, dataSizeForCheck = {}," +
+ " dataThreshold = {}", persistenceId(), journalSize, dataSizeForCheck, dataThreshold);
+
long lastAppliedIndex = -1;
long lastAppliedTerm = -1;
ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied());
- if (lastAppliedEntry != null) {
+ if (!hasFollowers()) {
+ lastAppliedIndex = replicatedLogEntry.getIndex();
+ lastAppliedTerm = replicatedLogEntry.getTerm();
+ } else if (lastAppliedEntry != null) {
lastAppliedIndex = lastAppliedEntry.getIndex();
lastAppliedTerm = lastAppliedEntry.getTerm();
}
- LOG.debug("Snapshot Capture logSize: {}", journal.size());
- LOG.debug("Snapshot Capture lastApplied:{} ", context.getLastApplied());
- LOG.debug("Snapshot Capture lastAppliedIndex:{}", lastAppliedIndex);
- LOG.debug("Snapshot Capture lastAppliedTerm:{}", lastAppliedTerm);
-
- // create a snapshot object from the state provided and save it
- // when snapshot is saved async, SaveSnapshotSuccess is raised.
- Snapshot sn = Snapshot.create(createSnapshot(),
- getFrom(context.getLastApplied() + 1),
- lastIndex(), lastTerm(), lastAppliedIndex,
- lastAppliedTerm);
- saveSnapshot(sn);
-
- LOG.info("Persisting of snapshot done:{}", sn.getLogMessage());
-
- //be greedy and remove entries from in-mem journal which are in the snapshot
- // and update snapshotIndex and snapshotTerm without waiting for the success,
- // TODO: damage-recovery to be done on failure
- journal.subList(0, (int) (lastAppliedIndex - snapshotIndex)).clear();
- snapshotIndex = lastAppliedIndex;
- snapshotTerm = lastAppliedTerm;
-
- LOG.info("Removed in-memory snapshotted entries, " +
- "adjusted snaphsotIndex:{}" +
- "and term:{}", snapshotIndex, lastAppliedTerm);
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Snapshot Capture logSize: {}", persistenceId(), journal.size());
+ LOG.debug("{}: Snapshot Capture lastApplied:{} ",
+ persistenceId(), context.getLastApplied());
+ LOG.debug("{}: Snapshot Capture lastAppliedIndex:{}", persistenceId(),
+ lastAppliedIndex);
+ LOG.debug("{}: Snapshot Capture lastAppliedTerm:{}", persistenceId(),
+ lastAppliedTerm);
+ }
+
+ // send a CaptureSnapshot to self to make the expensive operation async.
+ long replicatedToAllIndex = getCurrentBehavior().getReplicatedToAllIndex();
+ ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
+ getSelf().tell(new CaptureSnapshot(lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
+ (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
+ (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1)),
+ null);
+ context.setSnapshotCaptureInitiated(true);
}
- // Send message for replication
- if (clientActor != null) {
- currentBehavior.handleMessage(getSelf(),
- new Replicate(clientActor, identifier,
- replicatedLogEntry)
- );
+ if (callback != null){
+ callback.apply(replicatedLogEntry);
}
}
}
@@ -528,10 +842,10 @@ public abstract class RaftActor extends UntypedPersistentActor {
}
- private static class DeleteEntries implements Serializable {
+ static class DeleteEntries implements Serializable {
+ private static final long serialVersionUID = 1L;
private final int fromIndex;
-
public DeleteEntries(int fromIndex) {
this.fromIndex = fromIndex;
}
@@ -542,65 +856,6 @@ public abstract class RaftActor extends UntypedPersistentActor {
}
- private static class Snapshot implements Serializable {
- private final Object state;
- private final List unAppliedEntries;
- private final long lastIndex;
- private final long lastTerm;
- private final long lastAppliedIndex;
- private final long lastAppliedTerm;
-
- private Snapshot(Object state,
- List unAppliedEntries, long lastIndex,
- long lastTerm, long lastAppliedIndex, long lastAppliedTerm) {
- this.state = state;
- this.unAppliedEntries = unAppliedEntries;
- this.lastIndex = lastIndex;
- this.lastTerm = lastTerm;
- this.lastAppliedIndex = lastAppliedIndex;
- this.lastAppliedTerm = lastAppliedTerm;
- }
-
-
- public static Snapshot create(Object state,
- List entries, long lastIndex, long lastTerm,
- long lastAppliedIndex, long lastAppliedTerm) {
- return new Snapshot(state, entries, lastIndex, lastTerm,
- lastAppliedIndex, lastAppliedTerm);
- }
-
- public Object getState() {
- return state;
- }
-
- public List getUnAppliedEntries() {
- return unAppliedEntries;
- }
-
- public long getLastTerm() {
- return lastTerm;
- }
-
- public long getLastAppliedIndex() {
- return lastAppliedIndex;
- }
-
- public long getLastAppliedTerm() {
- return lastAppliedTerm;
- }
-
- public String getLogMessage() {
- StringBuilder sb = new StringBuilder();
- return sb.append("Snapshot={")
- .append("lastTerm:" + this.getLastTerm() + ", ")
- .append("LastAppliedIndex:" + this.getLastAppliedIndex() + ", ")
- .append("LastAppliedTerm:" + this.getLastAppliedTerm() + ", ")
- .append("UnAppliedEntries size:" + this.getUnAppliedEntries().size() + "}")
- .toString();
-
- }
- }
-
private class ElectionTermImpl implements ElectionTerm {
/**
* Identifier of the actor whose election term information this is
@@ -608,17 +863,20 @@ public abstract class RaftActor extends UntypedPersistentActor {
private long currentTerm = 0;
private String votedFor = null;
+ @Override
public long getCurrentTerm() {
return currentTerm;
}
+ @Override
public String getVotedFor() {
return votedFor;
}
@Override public void update(long currentTerm, String votedFor) {
- LOG.debug("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
-
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
+ }
this.currentTerm = currentTerm;
this.votedFor = votedFor;
}
@@ -627,7 +885,7 @@ public abstract class RaftActor extends UntypedPersistentActor {
public void updateAndPersist(long currentTerm, String votedFor){
update(currentTerm, votedFor);
// FIXME : Maybe first persist then update the state
- persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure(){
+ persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure(){
@Override public void apply(UpdateElectionTerm param)
throws Exception {
@@ -637,7 +895,8 @@ public abstract class RaftActor extends UntypedPersistentActor {
}
}
- private static class UpdateElectionTerm implements Serializable {
+ static class UpdateElectionTerm implements Serializable {
+ private static final long serialVersionUID = 1L;
private final long currentTerm;
private final String votedFor;
@@ -655,4 +914,38 @@ public abstract class RaftActor extends UntypedPersistentActor {
}
}
+ protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider {
+
+ public NonPersistentRaftDataProvider(){
+
+ }
+
+ /**
+ * The way snapshotting works is,
+ *
+ * - RaftActor calls createSnapshot on the Shard
+ *
- Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot
+ *
- When saveSnapshot is invoked on the akka-persistence API it uses the SnapshotStore to save the snapshot.
+ * The SnapshotStore sends SaveSnapshotSuccess or SaveSnapshotFailure. When the RaftActor gets SaveSnapshot
+ * success it commits the snapshot to the in-memory journal. This commitSnapshot is mimicking what is done
+ * in SaveSnapshotSuccess.
+ *
+ * @param o
+ */
+ @Override
+ public void saveSnapshot(Object o) {
+ // Make saving Snapshot successful
+ commitSnapshot(-1L);
+ }
+ }
+
+ @VisibleForTesting
+ void setCurrentBehavior(AbstractRaftActorBehavior behavior) {
+ currentBehavior = behavior;
+ }
+
+ protected RaftActorBehavior getCurrentBehavior() {
+ return currentBehavior;
+ }
+
}