X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?p=controller.git;a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fsal-akka-raft%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fraft%2FRaftActor.java;h=41a807aa355d394f659f488209f65e732646b120;hp=eaa005e3efdf6911a7beadce66e427d68a89161e;hb=c31a6fcf9fb070d4419ca4c32d8b531fdcb5030d;hpb=d8d8f731bbe6c58fcbd0e616734e2e230aaf4ab4 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 eaa005e3ef..41a807aa35 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 @@ -11,34 +11,38 @@ package org.opendaylight.controller.cluster.raft; import akka.actor.ActorRef; import akka.actor.ActorSelection; 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 com.google.common.annotations.VisibleForTesting; +import com.google.common.base.Objects; import com.google.common.base.Optional; -import com.google.common.base.Stopwatch; -import com.google.protobuf.ByteString; +import com.google.common.collect.ImmutableMap; +import com.google.common.collect.Lists; import java.io.Serializable; +import java.util.Collection; +import java.util.List; import java.util.Map; import java.util.concurrent.TimeUnit; +import javax.annotation.Nonnull; +import org.apache.commons.lang3.time.DurationFormatUtils; import org.opendaylight.controller.cluster.DataPersistenceProvider; +import org.opendaylight.controller.cluster.DelegatingPersistentDataProvider; +import org.opendaylight.controller.cluster.NonPersistentDataProvider; +import org.opendaylight.controller.cluster.PersistentDataProvider; import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor; +import org.opendaylight.controller.cluster.notifications.LeaderStateChanged; import org.opendaylight.controller.cluster.notifications.RoleChanged; import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries; -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 org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot; -import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior; +import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader; +import org.opendaylight.controller.cluster.raft.behaviors.DelegatingRaftActorBehavior; import org.opendaylight.controller.cluster.raft.behaviors.Follower; import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior; 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.FollowerInfo; +import org.opendaylight.controller.cluster.raft.client.messages.GetOnDemandRaftState; +import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState; import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload; import org.slf4j.Logger; import org.slf4j.LoggerFactory; @@ -87,20 +91,13 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { private static final long APPLY_STATE_DELAY_THRESHOLD_IN_NANOS = TimeUnit.MILLISECONDS.toNanos(50L); // 50 millis - private static final Procedure APPLY_JOURNAL_ENTRIES_PERSIST_CALLBACK = - new Procedure() { - @Override - public void apply(ApplyJournalEntries param) throws Exception { - } - }; - protected final Logger LOG = LoggerFactory.getLogger(getClass()); /** * The current state determines the current behavior of a RaftActor * A Raft Actor always starts off in the Follower State */ - private RaftActorBehavior currentBehavior; + private final DelegatingRaftActorBehavior currentBehavior = new DelegatingRaftActorBehavior(); /** * This context should NOT be passed directly to any other actor it is @@ -108,16 +105,13 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { */ private final RaftActorContextImpl context; - /** - * The in-memory journal - */ - private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl(); + private final DelegatingPersistentDataProvider delegatingPersistenceProvider = new DelegatingPersistentDataProvider(null); - private CaptureSnapshot captureSnapshot = null; + private RaftActorRecoverySupport raftRecovery; - private Stopwatch recoveryTimer; + private RaftActorSnapshotMessageSupport snapshotSupport; - private int currentRecoveryBatchCount; + private final BehaviorStateHolder reusableBehaviorStateHolder = new BehaviorStateHolder(); public RaftActor(String id, Map peerAddresses) { this(id, peerAddresses, Optional.absent()); @@ -127,16 +121,11 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { Optional configParams) { context = new RaftActorContextImpl(this.getSelf(), - this.getContext(), id, new ElectionTermImpl(), - -1, -1, replicatedLog, peerAddresses, - (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()), - LOG); - } + this.getContext(), id, new ElectionTermImpl(delegatingPersistenceProvider, id, LOG), + -1, -1, peerAddresses, + (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()), LOG); - private void initRecoveryTimer() { - if(recoveryTimer == null) { - recoveryTimer = Stopwatch.createStarted(); - } + context.setReplicatedLog(ReplicatedLogImpl.newInstance(context, delegatingPersistenceProvider, currentBehavior)); } @Override @@ -149,7 +138,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { @Override public void postStop() { - if(currentBehavior != null) { + if(currentBehavior.getDelegate() != null) { try { currentBehavior.close(); } catch (Exception e) { @@ -162,147 +151,52 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { @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) { - // Handle this message for backwards compatibility with pre-Lithium versions. - onRecoveredApplyLogEntries(((ApplyLogEntries) message).getToIndex()); - } else if (message instanceof ApplyJournalEntries) { - onRecoveredApplyLogEntries(((ApplyJournalEntries) message).getToIndex()); - } 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) { + if(raftRecovery == null) { + raftRecovery = new RaftActorRecoverySupport(delegatingPersistenceProvider, context, currentBehavior, + getRaftActorRecoveryCohort()); + } + + boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message); + if(recoveryComplete) { + if(!persistence().isRecoveryApplicable()) { // 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()); // Delete all the akka snapshots as they will not be needed deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue())); - - onRecoveryComplete(); - - 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()); - } + onRecoveryComplete(); - replicatedLog.append(logEntry); - } + initializeBehavior(); - private void onRecoveredApplyLogEntries(long toIndex) { - if(LOG.isDebugEnabled()) { - LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}", - persistenceId(), context.getLastApplied() + 1, toIndex); + raftRecovery = null; } - - for (long i = context.getLastApplied() + 1; i <= toIndex; i++) { - batchRecoveredLogEntry(replicatedLog.get(i)); - } - - context.setLastApplied(toIndex); - context.setCommitIndex(toIndex); } - private void batchRecoveredLogEntry(ReplicatedLogEntry logEntry) { - initRecoveryTimer(); - - int batchSize = context.getConfigParams().getJournalRecoveryLogBatchSize(); - if(currentRecoveryBatchCount == 0) { - startLogRecoveryBatch(batchSize); - } - - appendRecoveredLogEntry(logEntry.getData()); - - if(++currentRecoveryBatchCount >= batchSize) { - endCurrentLogRecoveryBatch(); - } + protected void initializeBehavior(){ + changeCurrentBehavior(new Follower(context)); } - private void endCurrentLogRecoveryBatch() { - applyCurrentLogRecoveryBatch(); - currentRecoveryBatchCount = 0; + protected void changeCurrentBehavior(RaftActorBehavior newBehavior){ + reusableBehaviorStateHolder.init(getCurrentBehavior()); + setCurrentBehavior(newBehavior); + handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior()); } - private void onRecoveryCompletedMessage() { - if(currentRecoveryBatchCount > 0) { - endCurrentLogRecoveryBatch(); + @Override + public void handleCommand(Object message) { + if(snapshotSupport == null) { + snapshotSupport = new RaftActorSnapshotMessageSupport(delegatingPersistenceProvider, context, + currentBehavior, getRaftActorSnapshotCohort(), self()); } - onRecoveryComplete(); - - String recoveryTime = ""; - if(recoveryTimer != null) { - recoveryTimer.stop(); - recoveryTime = " in " + recoveryTimer.toString(); - recoveryTimer = null; + boolean handled = snapshotSupport.handleSnapshotMessage(message); + if(handled) { + return; } - 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(); - } - - 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; @@ -327,88 +221,91 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), applyEntries.getToIndex()); } - persistence().persist(applyEntries, APPLY_JOURNAL_ENTRIES_PERSIST_CALLBACK); - - } else if(message instanceof ApplySnapshot ) { - 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()); + persistence().persist(applyEntries, NoopProcedure.instance()); } else if (message instanceof FindLeader) { getSender().tell( new FindLeaderReply(getLeaderAddress()), getSelf() ); - - } else if (message instanceof SaveSnapshotSuccess) { - SaveSnapshotSuccess success = (SaveSnapshotSuccess) message; - LOG.info("{}: SaveSnapshotSuccess received for snapshot", persistenceId()); - - 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()); - - context.getReplicatedLog().snapshotRollback(); - - 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()); - - } else if (message instanceof CaptureSnapshot) { - LOG.debug("{}: CaptureSnapshot received by actor: {}", persistenceId(), message); - - if(captureSnapshot == null) { - captureSnapshot = (CaptureSnapshot)message; - createSnapshot(); + } else if(message instanceof GetOnDemandRaftState) { + onGetOnDemandRaftStats(); + } else { + reusableBehaviorStateHolder.init(getCurrentBehavior()); + + setCurrentBehavior(currentBehavior.handleMessage(getSender(), message)); + + handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior()); + } + } + + private void onGetOnDemandRaftStats() { + // Debugging message to retrieve raft stats. + + OnDemandRaftState.Builder builder = OnDemandRaftState.builder() + .commitIndex(context.getCommitIndex()) + .currentTerm(context.getTermInformation().getCurrentTerm()) + .inMemoryJournalDataSize(replicatedLog().dataSize()) + .inMemoryJournalLogSize(replicatedLog().size()) + .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing()) + .lastApplied(context.getLastApplied()) + .lastIndex(replicatedLog().lastIndex()) + .lastTerm(replicatedLog().lastTerm()) + .leader(getLeaderId()) + .raftState(currentBehavior.state().toString()) + .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex()) + .snapshotIndex(replicatedLog().getSnapshotIndex()) + .snapshotTerm(replicatedLog().getSnapshotTerm()) + .votedFor(context.getTermInformation().getVotedFor()) + .peerAddresses(ImmutableMap.copyOf(context.getPeerAddresses())); + + ReplicatedLogEntry lastLogEntry = getLastLogEntry(); + if (lastLogEntry != null) { + builder.lastLogIndex(lastLogEntry.getIndex()); + builder.lastLogTerm(lastLogEntry.getTerm()); + } + + if(getCurrentBehavior() instanceof AbstractLeader) { + AbstractLeader leader = (AbstractLeader)getCurrentBehavior(); + Collection followerIds = leader.getFollowerIds(); + List followerInfoList = Lists.newArrayListWithCapacity(followerIds.size()); + for(String id: followerIds) { + final FollowerLogInformation info = leader.getFollower(id); + followerInfoList.add(new FollowerInfo(id, info.getNextIndex(), info.getMatchIndex(), + info.isFollowerActive(), DurationFormatUtils.formatDurationHMS(info.timeSinceLastActivity()))); } - } else if (message instanceof CaptureSnapshotReply){ - handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot()); + builder.followerInfoList(followerInfoList); + } - } else { - RaftActorBehavior oldBehavior = currentBehavior; - currentBehavior = currentBehavior.handleMessage(getSender(), message); + sender().tell(builder.build(), self()); - handleBehaviorChange(oldBehavior, currentBehavior); - } } - private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) { + private void handleBehaviorChange(BehaviorStateHolder oldBehaviorState, RaftActorBehavior currentBehavior) { + RaftActorBehavior oldBehavior = oldBehaviorState.getBehavior(); + if (oldBehavior != currentBehavior){ onStateChanged(); } - String oldBehaviorLeaderId = oldBehavior == null? null : oldBehavior.getLeaderId(); - String oldBehaviorState = oldBehavior == null? null : oldBehavior.state().name(); + String oldBehaviorLeaderId = oldBehavior == null ? null : oldBehaviorState.getLeaderId(); + String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name(); // it can happen that the state has not changed but the leader has changed. - onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId()); + Optional roleChangeNotifier = getRoleChangeNotifier(); + if(!Objects.equal(oldBehaviorLeaderId, currentBehavior.getLeaderId())) { + if(roleChangeNotifier.isPresent()) { + roleChangeNotifier.get().tell(new LeaderStateChanged(getId(), currentBehavior.getLeaderId()), getSelf()); + } + + onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId()); + } - if (getRoleChangeNotifier().isPresent() && + if (roleChangeNotifier.isPresent() && (oldBehavior == null || (oldBehavior.state() != currentBehavior.state()))) { - getRoleChangeNotifier().get().tell( - new RoleChanged(getId(), oldBehaviorState , currentBehavior.state().name()), - getSelf()); + roleChangeNotifier.get().tell(new RoleChanged(getId(), oldBehaviorStateName , + currentBehavior.state().name()), getSelf()); } } @@ -433,46 +330,49 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { final RaftActorContext raftContext = getRaftActorContext(); - replicatedLog - .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 ApplyJournalEntries message so that we write the fact that we applied - // the state to durable storage - self().tell(new ApplyJournalEntries(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) - ); - } - - } - }); } + replicatedLog().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 ApplyJournalEntries message so that we write the fact that we applied + // the state to durable storage + self().tell(new ApplyJournalEntries(replicatedLogEntry.getIndex()), self()); + + context.getSnapshotManager().trimLog(context.getLastApplied(), currentBehavior); + + } else if (clientActor != null) { + // Send message for replication + currentBehavior.handleMessage(getSelf(), + new Replicate(clientActor, identifier, replicatedLogEntry)); + } + } + }); + } + + private ReplicatedLog replicatedLog() { + return context.getReplicatedLog(); + } protected String getId() { return context.getId(); } + @VisibleForTesting + void setCurrentBehavior(RaftActorBehavior behavior) { + currentBehavior.setDelegate(behavior); + } + + protected RaftActorBehavior getCurrentBehavior() { + return currentBehavior.getDelegate(); + } + /** * Derived actors can call the isLeader method to check if the current * RaftActor is the Leader or not @@ -513,7 +413,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected ReplicatedLogEntry getLastLogEntry() { - return replicatedLog.last(); + return replicatedLog().last(); } protected Long getCurrentTerm(){ @@ -536,6 +436,41 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { context.setConfigParams(configParams); } + public final DataPersistenceProvider persistence() { + return delegatingPersistenceProvider.getDelegate(); + } + + public void setPersistence(DataPersistenceProvider provider) { + delegatingPersistenceProvider.setDelegate(provider); + } + + protected void setPersistence(boolean persistent) { + if(persistent) { + setPersistence(new PersistentDataProvider(this)); + } else { + setPersistence(new NonPersistentDataProvider() { + /** + * The way snapshotting works is, + *
    + *
  1. RaftActor calls createSnapshot on the Shard + *
  2. Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot + *
  3. 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. + *
+ */ + @Override + public void saveSnapshot(Object o) { + // Make saving Snapshot successful + // Committing the snapshot here would end up calling commit in the creating state which would + // be a state violation. That's why now we send a message to commit the snapshot. + self().tell(RaftActorSnapshotMessageSupport.COMMIT_SNAPSHOT, self()); + } + }); + } + } + /** * setPeerAddress sets the address of a known peer at a later time. *

@@ -554,13 +489,6 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { 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 * needs to be applied to the actor's state @@ -582,31 +510,10 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { 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}. + * Returns the RaftActorRecoveryCohort to participate in persistence recovery. */ - protected abstract void applyCurrentLogRecoveryBatch(); + @Nonnull + protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort(); /** * This method is called when recovery is complete. @@ -614,24 +521,10 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { 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. - *

- * During recovery the state that is returned by the derived actor will - * be passed back to it by calling the applySnapshot method - * - * @return The current state of the actor + * Returns the RaftActorSnapshotCohort to participate in persistence recovery. */ - protected abstract void createSnapshot(); - - /** - * 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 snapshotBytes A snapshot of the state of the actor - */ - protected abstract void applySnapshot(byte[] snapshotBytes); + @Nonnull + protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort(); /** * This method will be called by the RaftActor when the state of the @@ -640,8 +533,6 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { */ protected abstract void onStateChanged(); - 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. @@ -650,17 +541,6 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { 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. - persistence().deleteSnapshots(new SnapshotSelectionCriteria( - sequenceNumber - context.getConfigParams().getSnapshotBatchCount(), 43200000)); - - // Trim akka journal - persistence().deleteMessages(sequenceNumber); - } - private String getLeaderAddress(){ if(isLeader()){ return getSelf().path().toString(); @@ -678,214 +558,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { return peerAddress; } - private void handleCaptureSnapshotReply(byte[] snapshotBytes) { - LOG.debug("{}: 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(LOG.isDebugEnabled()) { - LOG.debug("{}: dataSize {} exceeds dataThreshold {} - doing snapshotPreCommit with index {}", - persistenceId(), context.getReplicatedLog().dataSize(), dataThreshold, - captureSnapshot.getLastAppliedIndex()); - } - - // 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()); - - // Don't reset replicatedToAllIndex to -1 as this may prevent us from trimming the log after an - // install snapshot to a follower. - if(captureSnapshot.getReplicatedToAllIndex() >= 0) { - getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex()); - } - } else if(captureSnapshot.getReplicatedToAllIndex() != -1){ - // clear the log based on replicatedToAllIndex - context.getReplicatedLog().snapshotPreCommit(captureSnapshot.getReplicatedToAllIndex(), - captureSnapshot.getReplicatedToAllTerm()); - - getCurrentBehavior().setReplicatedToAllIndex(captureSnapshot.getReplicatedToAllIndex()); - } else { - // The replicatedToAllIndex was not found in the log - // This means that replicatedToAllIndex never moved beyond -1 or that it is already in the snapshot. - // In this scenario we may need to save the snapshot to the akka persistence - // snapshot for recovery but we do not need to do the replicated log trimming. - context.getReplicatedLog().snapshotPreCommit(replicatedLog.getSnapshotIndex(), - replicatedLog.getSnapshotTerm()); - } - - - LOG.info("{}: Removed in-memory snapshotted entries, adjusted snaphsotIndex: {} " + - "and term: {}", persistenceId(), replicatedLog.getSnapshotIndex(), - replicatedLog.getSnapshotTerm()); - - 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.getLastAppliedIndex(), snapshot.getLastAppliedTerm(), - snapshot.getUnAppliedEntries()); - } - - public ReplicatedLogImpl() { - super(); - } - - @Override public void removeFromAndPersist(long logEntryIndex) { - int adjustedIndex = adjustedIndex(logEntryIndex); - - if (adjustedIndex < 0) { - return; - } - - // FIXME: Maybe this should be done after the command is saved - journal.subList(adjustedIndex , journal.size()).clear(); - - persistence().persist(new DeleteEntries(adjustedIndex), new Procedure() { - - @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(replicatedLogEntry, null); - } - - 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); - - // When persisting events with persist it is guaranteed that the - // persistent actor will not receive further commands between the - // persist call and the execution(s) of the associated event - // handler. This also holds for multiple persist calls in context - // of a single command. - persistence().persist(replicatedLogEntry, - new Procedure() { - @Override - public void apply(ReplicatedLogEntry evt) throws Exception { - 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 (!hasFollowers()) { - lastAppliedIndex = replicatedLogEntry.getIndex(); - lastAppliedTerm = replicatedLogEntry.getTerm(); - } else if (lastAppliedEntry != null) { - lastAppliedIndex = lastAppliedEntry.getIndex(); - lastAppliedTerm = lastAppliedEntry.getTerm(); - } - - 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); - } - if (callback != null){ - callback.apply(replicatedLogEntry); - } - } - } - ); - } - + return getRaftActorContext().hasFollowers(); } static class DeleteEntries implements Serializable { @@ -901,46 +575,6 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } } - - private class ElectionTermImpl implements ElectionTerm { - /** - * Identifier of the actor whose election term information this is - */ - 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) { - if(LOG.isDebugEnabled()) { - LOG.debug("{}: Set currentTerm={}, votedFor={}", persistenceId(), currentTerm, votedFor); - } - this.currentTerm = currentTerm; - this.votedFor = votedFor; - } - - @Override - public void updateAndPersist(long currentTerm, String votedFor){ - update(currentTerm, votedFor); - // FIXME : Maybe first persist then update the state - persistence().persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure(){ - - @Override public void apply(UpdateElectionTerm param) - throws Exception { - - } - }); - } - } - static class UpdateElectionTerm implements Serializable { private static final long serialVersionUID = 1L; private final long currentTerm; @@ -960,38 +594,21 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } } - protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider { - - public NonPersistentRaftDataProvider(){ + private static class BehaviorStateHolder { + private RaftActorBehavior behavior; + private String leaderId; + void init(RaftActorBehavior behavior) { + this.behavior = behavior; + this.leaderId = behavior != null ? behavior.getLeaderId() : null; } - /** - * The way snapshotting works is, - *

    - *
  1. RaftActor calls createSnapshot on the Shard - *
  2. Shard sends a CaptureSnapshotReply and RaftActor then calls saveSnapshot - *
  3. 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); + RaftActorBehavior getBehavior() { + return behavior; } - } - @VisibleForTesting - void setCurrentBehavior(AbstractRaftActorBehavior behavior) { - currentBehavior = behavior; - } - - protected RaftActorBehavior getCurrentBehavior() { - return currentBehavior; + String getLeaderId() { + return leaderId; + } } - }