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=b93ea4dd8a8b3e6028d2943f87c84d17ee3eb01a;hp=b74259d4851153659df0c2866f6323b9234eff06;hb=8882e6077db69d22bcc57fcf12dd4a02a81a4967;hpb=edcc020c8fda4b13f22a31d79c13feef0b53b0ee 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 b74259d485..b93ea4dd8a 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 @@ -1,5 +1,6 @@ /* * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved. + * Copyright (c) 2015 Brocade Communications Systems, Inc. and others. All rights reserved. * * This program and the accompanying materials are made available under the * terms of the Eclipse Public License v1.0 which accompanies this distribution, @@ -11,36 +12,33 @@ 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.common.collect.ImmutableMap; +import com.google.common.base.Supplier; import com.google.common.collect.Lists; import java.io.Serializable; import java.util.Collection; +import java.util.HashMap; 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.InitiateCaptureSnapshot; import org.opendaylight.controller.cluster.raft.base.messages.Replicate; +import org.opendaylight.controller.cluster.raft.base.messages.SwitchBehavior; import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader; -import org.opendaylight.controller.cluster.raft.behaviors.AbstractRaftActorBehavior; +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; @@ -96,21 +94,13 @@ 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 { - } - }; - private static final String COMMIT_SNAPSHOT = "commit_snapshot"; - 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 @@ -118,37 +108,34 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { */ private final RaftActorContextImpl context; - private final Procedure createSnapshotProcedure = new CreateSnapshotProcedure(); + private final DelegatingPersistentDataProvider delegatingPersistenceProvider; - /** - * The in-memory journal - */ - private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl(); + private final PersistentDataProvider persistentProvider; - private Stopwatch recoveryTimer; + private RaftActorRecoverySupport raftRecovery; - private int currentRecoveryBatchCount; + private RaftActorSnapshotMessageSupport snapshotSupport; private final BehaviorStateHolder reusableBehaviorStateHolder = new BehaviorStateHolder(); - public RaftActor(String id, Map peerAddresses) { - this(id, peerAddresses, Optional.absent()); - } + private final SwitchBehaviorSupplier reusableSwitchBehaviorSupplier = new SwitchBehaviorSupplier(); + + private RaftActorServerConfigurationSupport serverConfigurationSupport; public RaftActor(String id, Map peerAddresses, - Optional configParams) { + Optional configParams, short payloadVersion) { + + persistentProvider = new PersistentDataProvider(this); + delegatingPersistenceProvider = new RaftActorDelegatingPersistentDataProvider(null, persistentProvider); context = new RaftActorContextImpl(this.getSelf(), - this.getContext(), id, new ElectionTermImpl(), - -1, -1, replicatedLog, peerAddresses, + this.getContext(), id, new ElectionTermImpl(persistentProvider, id, LOG), + -1, -1, peerAddresses, (configParams.isPresent() ? configParams.get(): new DefaultConfigParamsImpl()), - LOG); - } + delegatingPersistenceProvider, LOG); - private void initRecoveryTimer() { - if(recoveryTimer == null) { - recoveryTimer = Stopwatch.createStarted(); - } + context.setPayloadVersion(payloadVersion); + context.setReplicatedLog(ReplicatedLogImpl.newInstance(context, currentBehavior)); } @Override @@ -157,11 +144,14 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { context.getConfigParams().getJournalRecoveryLogBatchSize()); super.preStart(); + + snapshotSupport = newRaftActorSnapshotMessageSupport(); + serverConfigurationSupport = new RaftActorServerConfigurationSupport(getRaftActorContext()); } @Override public void postStop() { - if(currentBehavior != null) { + if(currentBehavior.getDelegate() != null) { try { currentBehavior.close(); } catch (Exception e) { @@ -174,134 +164,29 @@ 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) { - // 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()); + if(raftRecovery == null) { + raftRecovery = newRaftActorRecoverySupport(); } - 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()); + boolean recoveryComplete = raftRecovery.handleRecoveryMessage(message, persistentProvider); + if(recoveryComplete) { + onRecoveryComplete(); - Stopwatch timer = Stopwatch.createStarted(); + initializeBehavior(); - // Apply the snapshot to the actors state - applyRecoverySnapshot(snapshot.getState()); + raftRecovery = null; - 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(long toIndex) { - if(LOG.isDebugEnabled()) { - LOG.debug("{}: Received ApplyLogEntries for recovery, applying to state: {} to {}", - persistenceId(), context.getLastApplied() + 1, toIndex); - } - - 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(); + if (context.getReplicatedLog().size() > 0) { + self().tell(new InitiateCaptureSnapshot(), self()); + LOG.info("Snapshot capture initiated after recovery"); + } else { + LOG.info("Snapshot capture NOT initiated after recovery, journal empty"); + } } } - 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(); + protected RaftActorRecoverySupport newRaftActorRecoverySupport() { + return new RaftActorRecoverySupport(context, currentBehavior, getRaftActorRecoveryCohort()); } protected void initializeBehavior(){ @@ -309,13 +194,16 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected void changeCurrentBehavior(RaftActorBehavior newBehavior){ - reusableBehaviorStateHolder.init(currentBehavior); - currentBehavior = newBehavior; - handleBehaviorChange(reusableBehaviorStateHolder, currentBehavior); + reusableBehaviorStateHolder.init(getCurrentBehavior()); + setCurrentBehavior(newBehavior); + handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior()); } - @Override public void handleCommand(Object message) { - if (message instanceof ApplyState){ + @Override + public void handleCommand(final Object message) { + if(serverConfigurationSupport.handleMessage(message, this, getSender())) { + return; + } else if (message instanceof ApplyState){ ApplyState applyState = (ApplyState) message; long elapsedTime = (System.nanoTime() - applyState.getStartTime()); @@ -333,92 +221,89 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { applyState(applyState.getClientActor(), applyState.getIdentifier(), applyState.getReplicatedLogEntry().getData()); + if (!hasFollowers()) { + // for single node, the capture should happen after the apply state + // as we delete messages from the persistent journal which have made it to the snapshot + // capturing the snapshot before applying makes the persistent journal and snapshot out of sync + // and recovery shows data missing + context.getReplicatedLog().captureSnapshotIfReady(applyState.getReplicatedLogEntry()); + + context.getSnapshotManager().trimLog(context.getLastApplied(), currentBehavior); + } + } else if (message instanceof ApplyJournalEntries){ ApplyJournalEntries applyEntries = (ApplyJournalEntries) message; if(LOG.isDebugEnabled()) { 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.getSnapshotManager().rollback(); - - } else if (message instanceof CaptureSnapshot) { - LOG.debug("{}: CaptureSnapshot received by actor: {}", persistenceId(), message); - - context.getSnapshotManager().create(createSnapshotProcedure); - - } else if (message instanceof CaptureSnapshotReply) { - handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot()); } else if(message instanceof GetOnDemandRaftState) { onGetOnDemandRaftStats(); - } else if (message.equals(COMMIT_SNAPSHOT)) { - commitSnapshot(-1); - } else { - reusableBehaviorStateHolder.init(currentBehavior); + } else if(message instanceof InitiateCaptureSnapshot) { + captureSnapshot(); + } else if(message instanceof SwitchBehavior){ + switchBehavior(((SwitchBehavior) message)); + } else if(!snapshotSupport.handleSnapshotMessage(message, getSender())) { + switchBehavior(reusableSwitchBehaviorSupplier.handleMessage(getSender(), message)); + } + } + + private void switchBehavior(SwitchBehavior message) { + if(!getRaftActorContext().getRaftPolicy().automaticElectionsEnabled()) { + RaftState newState = message.getNewState(); + if( newState == RaftState.Leader || newState == RaftState.Follower) { + switchBehavior(reusableSwitchBehaviorSupplier.handleMessage(getSender(), message)); + getRaftActorContext().getTermInformation().updateAndPersist(message.getNewTerm(), ""); + } else { + LOG.warn("Switching to behavior : {} - not supported", newState); + } + } + } - currentBehavior = currentBehavior.handleMessage(getSender(), message); + private void switchBehavior(Supplier supplier){ + reusableBehaviorStateHolder.init(getCurrentBehavior()); - handleBehaviorChange(reusableBehaviorStateHolder, currentBehavior); - } + setCurrentBehavior(supplier.get()); + + handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior()); + } + + protected RaftActorSnapshotMessageSupport newRaftActorSnapshotMessageSupport() { + return new RaftActorSnapshotMessageSupport(context, currentBehavior, + getRaftActorSnapshotCohort()); } private void onGetOnDemandRaftStats() { // Debugging message to retrieve raft stats. + Map peerAddresses = new HashMap<>(); + for(String peerId: context.getPeerIds()) { + peerAddresses.put(peerId, context.getPeerAddress(peerId)); + } + OnDemandRaftState.Builder builder = OnDemandRaftState.builder() .commitIndex(context.getCommitIndex()) .currentTerm(context.getTermInformation().getCurrentTerm()) - .inMemoryJournalDataSize(replicatedLog.dataSize()) - .inMemoryJournalLogSize(replicatedLog.size()) + .inMemoryJournalDataSize(replicatedLog().dataSize()) + .inMemoryJournalLogSize(replicatedLog().size()) .isSnapshotCaptureInitiated(context.getSnapshotManager().isCapturing()) .lastApplied(context.getLastApplied()) - .lastIndex(replicatedLog.lastIndex()) - .lastTerm(replicatedLog.lastTerm()) + .lastIndex(replicatedLog().lastIndex()) + .lastTerm(replicatedLog().lastTerm()) .leader(getLeaderId()) .raftState(currentBehavior.state().toString()) .replicatedToAllIndex(currentBehavior.getReplicatedToAllIndex()) - .snapshotIndex(replicatedLog.getSnapshotIndex()) - .snapshotTerm(replicatedLog.getSnapshotTerm()) + .snapshotIndex(replicatedLog().getSnapshotIndex()) + .snapshotTerm(replicatedLog().getSnapshotTerm()) .votedFor(context.getTermInformation().getVotedFor()) - .peerAddresses(ImmutableMap.copyOf(context.getPeerAddresses())); + .peerAddresses(peerAddresses); ReplicatedLogEntry lastLogEntry = getLastLogEntry(); if (lastLogEntry != null) { @@ -426,8 +311,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { builder.lastLogTerm(lastLogEntry.getTerm()); } - if(currentBehavior instanceof AbstractLeader) { - AbstractLeader leader = (AbstractLeader)currentBehavior; + if(getCurrentBehavior() instanceof AbstractLeader) { + AbstractLeader leader = (AbstractLeader)getCurrentBehavior(); Collection followerIds = leader.getFollowerIds(); List followerInfoList = Lists.newArrayListWithCapacity(followerIds.size()); for(String id: followerIds) { @@ -450,17 +335,19 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { onStateChanged(); } - String oldBehaviorLeaderId = oldBehavior == null ? null : oldBehaviorState.getLeaderId(); + String lastValidLeaderId = oldBehavior == null ? null : oldBehaviorState.getLastValidLeaderId(); String oldBehaviorStateName = oldBehavior == null ? null : oldBehavior.state().name(); // it can happen that the state has not changed but the leader has changed. Optional roleChangeNotifier = getRoleChangeNotifier(); - if(!Objects.equal(oldBehaviorLeaderId, currentBehavior.getLeaderId())) { + if(!Objects.equal(lastValidLeaderId, currentBehavior.getLeaderId()) || + oldBehaviorState.getLeaderPayloadVersion() != currentBehavior.getLeaderPayloadVersion()) { if(roleChangeNotifier.isPresent()) { - roleChangeNotifier.get().tell(new LeaderStateChanged(getId(), currentBehavior.getLeaderId()), getSelf()); + roleChangeNotifier.get().tell(newLeaderStateChanged(getId(), currentBehavior.getLeaderId(), + currentBehavior.getLeaderPayloadVersion()), getSelf()); } - onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId()); + onLeaderChanged(lastValidLeaderId, currentBehavior.getLeaderId()); } if (roleChangeNotifier.isPresent() && @@ -470,6 +357,20 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } } + protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) { + return new LeaderStateChanged(memberId, leaderId, leaderPayloadVersion); + } + + @Override + public long snapshotSequenceNr() { + // When we do a snapshot capture, we also capture and save the sequence-number of the persistent journal, + // so that we can delete the persistent journal based on the saved sequence-number + // However , when akka replays the journal during recovery, it replays it from the sequence number when the snapshot + // was saved and not the number we saved. + // We would want to override it , by asking akka to use the last-sequence number known to us. + return context.getSnapshotManager().getLastSequenceNumber(); + } + /** * When a derived RaftActor needs to persist something it must call * persistData. @@ -491,39 +392,49 @@ 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()); + replicatedLog().appendAndPersist(replicatedLogEntry, new Procedure() { + @Override + public void apply(ReplicatedLogEntry replicatedLogEntry) { + 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); + // Apply the state immediately. + self().tell(new ApplyState(clientActor, identifier, replicatedLogEntry), self()); - // 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()); + // 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) { + context.getReplicatedLog().captureSnapshotIfReady(replicatedLogEntry); - } else if (clientActor != null) { - // Send message for replication - currentBehavior.handleMessage(getSelf(), - new Replicate(clientActor, identifier, - replicatedLogEntry) - ); - } + // 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 @@ -564,7 +475,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected ReplicatedLogEntry getLastLogEntry() { - return replicatedLog.last(); + return replicatedLog().last(); } protected Long getCurrentTerm(){ @@ -584,7 +495,66 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected void updateConfigParams(ConfigParams configParams) { + + // obtain the RaftPolicy for oldConfigParams and the updated one. + String oldRaftPolicy = context.getConfigParams(). + getCustomRaftPolicyImplementationClass(); + String newRaftPolicy = configParams. + getCustomRaftPolicyImplementationClass(); + + LOG.debug("{}: RaftPolicy used with prev.config {}, RaftPolicy used with newConfig {}", persistenceId(), + oldRaftPolicy, newRaftPolicy); context.setConfigParams(configParams); + if (!Objects.equal(oldRaftPolicy, newRaftPolicy)) { + // The RaftPolicy was modified. If the current behavior is Follower then re-initialize to Follower + // but transfer the previous leaderId so it doesn't immediately try to schedule an election. This + // avoids potential disruption. Otherwise, switch to Follower normally. + RaftActorBehavior behavior = currentBehavior.getDelegate(); + if(behavior instanceof Follower) { + String previousLeaderId = ((Follower)behavior).getLeaderId(); + + LOG.debug("{}: Re-initializing to Follower with previous leaderId {}", persistenceId(), previousLeaderId); + + changeCurrentBehavior(new Follower(context, previousLeaderId)); + } else { + initializeBehavior(); + } + } + } + + 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()); + } + }); + } } /** @@ -605,10 +575,6 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { context.setPeerAddress(peerId, peerAddress); } - protected void commitSnapshot(long sequenceNumber) { - context.getSnapshotManager().commit(persistence(), sequenceNumber); - } - /** * The applyState method will be called by the RaftActor when some data * needs to be applied to the actor's state @@ -630,31 +596,10 @@ 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 + * Returns the RaftActorRecoveryCohort to participate in persistence recovery. */ - 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(); + @Nonnull + protected abstract RaftActorRecoveryCohort getRaftActorRecoveryCohort(); /** * This method is called when recovery is complete. @@ -662,24 +607,10 @@ 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 - */ - 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 + * Returns the RaftActorSnapshotCohort to participate in persistence recovery. */ - protected abstract void applySnapshot(byte[] snapshotBytes); + @Nonnull + protected abstract RaftActorSnapshotCohort getRaftActorSnapshotCohort(); /** * This method will be called by the RaftActor when the state of the @@ -688,8 +619,6 @@ 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. @@ -715,132 +644,30 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { return peerAddress; } - private void handleCaptureSnapshotReply(byte[] snapshotBytes) { - LOG.debug("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length); - - context.getSnapshotManager().persist(persistence(), snapshotBytes, currentBehavior, getTotalMemory()); - } - - protected long getTotalMemory() { - return Runtime.getRuntime().totalMemory(); - } - protected boolean hasFollowers(){ - return getRaftActorContext().getPeerAddresses().keySet().size() > 0; + return getRaftActorContext().hasFollowers(); } - private class ReplicatedLogImpl extends AbstractReplicatedLogImpl { - private static final int DATA_SIZE_DIVIDER = 5; - private long dataSizeSinceLastSnapshot = 0L; + private void captureSnapshot() { + SnapshotManager snapshotManager = context.getSnapshotManager(); + if(!snapshotManager.isCapturing()) { + LOG.debug("Take a snapshot of current state. lastReplicatedLog is {} and replicatedToAllIndex is {}", + replicatedLog().last(), currentBehavior.getReplicatedToAllIndex()); - public ReplicatedLogImpl(Snapshot snapshot) { - super(snapshot.getLastAppliedIndex(), snapshot.getLastAppliedTerm(), - snapshot.getUnAppliedEntries()); + snapshotManager.capture(replicatedLog().last(), currentBehavior.getReplicatedToAllIndex()); } - - 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; - - 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 journalSize = replicatedLogEntry.getIndex() + 1; - long dataThreshold = getTotalMemory() * - context.getConfigParams().getSnapshotDataThresholdPercentage() / 100; - - if ((journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 - || dataSizeForCheck > dataThreshold)) { - - boolean started = context.getSnapshotManager().capture(replicatedLogEntry, - currentBehavior.getReplicatedToAllIndex()); - - if(started){ - dataSizeSinceLastSnapshot = 0; - } - - } - - if (callback != null){ - callback.apply(replicatedLogEntry); - } - } - } - ); - } - } + /** + * @deprecated Deprecated in favor of {@link org.opendaylight.controller.cluster.raft.base.messages.DeleteEntries} + * whose type for fromIndex is long instead of int. This class was kept for backwards + * compatibility with Helium. + */ + // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility. + @SuppressWarnings("serial") + @Deprecated static class DeleteEntries implements Serializable { - private static final long serialVersionUID = 1L; private final int fromIndex; public DeleteEntries(int fromIndex) { @@ -852,48 +679,14 @@ 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 { - - } - }); - } - } - + /** + * @deprecated Deprecated in favor of non-inner class {@link org.opendaylight.controller.cluster.raft.base.messages.UpdateElectionTerm} + * which has serialVersionUID set. This class was kept for backwards compatibility with Helium. + */ + // Suppressing this warning as we can't set serialVersionUID to maintain backwards compatibility. + @SuppressWarnings("serial") + @Deprecated static class UpdateElectionTerm implements Serializable { - private static final long serialVersionUID = 1L; private final long currentTerm; private final String votedFor; @@ -911,66 +704,50 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } } - protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider { + private static class BehaviorStateHolder { + private RaftActorBehavior behavior; + private String lastValidLeaderId; + private short leaderPayloadVersion; - public NonPersistentRaftDataProvider(){ + void init(RaftActorBehavior behavior) { + this.behavior = behavior; + this.leaderPayloadVersion = behavior != null ? behavior.getLeaderPayloadVersion() : -1; + String behaviorLeaderId = behavior != null ? behavior.getLeaderId() : null; + if(behaviorLeaderId != null) { + this.lastValidLeaderId = behaviorLeaderId; + } } - /** - * 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 - // 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(COMMIT_SNAPSHOT, self()); + RaftActorBehavior getBehavior() { + return behavior; } - } - - private class CreateSnapshotProcedure implements Procedure { - - @Override - public void apply(Void aVoid) throws Exception { - createSnapshot(); + String getLastValidLeaderId() { + return lastValidLeaderId; } - } - @VisibleForTesting - void setCurrentBehavior(AbstractRaftActorBehavior behavior) { - currentBehavior = behavior; - } - - protected RaftActorBehavior getCurrentBehavior() { - return currentBehavior; + short getLeaderPayloadVersion() { + return leaderPayloadVersion; + } } - private static class BehaviorStateHolder { - private RaftActorBehavior behavior; - private String leaderId; + private class SwitchBehaviorSupplier implements Supplier { + private Object message; + private ActorRef sender; - void init(RaftActorBehavior behavior) { - this.behavior = behavior; - this.leaderId = behavior != null ? behavior.getLeaderId() : null; + public SwitchBehaviorSupplier handleMessage(ActorRef sender, Object message){ + this.sender = sender; + this.message = message; + return this; } - RaftActorBehavior getBehavior() { - return behavior; - } - - String getLeaderId() { - return leaderId; + @Override + public RaftActorBehavior get() { + if(this.message instanceof SwitchBehavior){ + return ((SwitchBehavior) message).getNewState().createBehavior(getRaftActorContext()); + } + return currentBehavior.handleMessage(sender, message); } } }