() {
+ @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();
@@ -269,18 +470,13 @@ public abstract class RaftActor extends UntypedPersistentActor {
* @return A reference to the leader if known, null otherwise
*/
protected ActorSelection getLeader(){
- String leaderId = currentBehavior.getLeaderId();
- if (leaderId == null) {
- return null;
- }
- String peerAddress = context.getPeerAddress(leaderId);
- LOG.debug("getLeader leaderId = " + leaderId + " peerAddress = "
- + peerAddress);
+ String leaderAddress = getLeaderAddress();
- if(peerAddress == null){
+ if(leaderAddress == null){
return null;
}
- return context.actorSelection(peerAddress);
+
+ return context.actorSelection(leaderAddress);
}
/**
@@ -311,6 +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.
*
@@ -329,7 +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
@@ -351,6 +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.
@@ -360,19 +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
@@ -381,55 +615,100 @@ 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(){
+ if(isLeader()){
+ return getSelf().path().toString();
+ }
+ String leaderId = currentBehavior.getLeaderId();
+ if (leaderId == null) {
+ return null;
+ }
+ String peerAddress = context.getPeerAddress(leaderId);
+ 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());
}
@@ -447,25 +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);
@@ -474,53 +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);
}
}
}
@@ -529,10 +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;
}
@@ -543,65 +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
@@ -609,17 +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.info("Set currentTerm={}, votedFor={}", currentTerm, votedFor);
-
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor);
+ }
this.currentTerm = currentTerm;
this.votedFor = votedFor;
}
@@ -628,7 +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 {
@@ -638,7 +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;
@@ -656,4 +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;
+ }
+
}