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=1c30fe23175b5af62ffe808d9428c3361688a4b0;hp=ec3f375bdeb0fb940e4d08d14397ebd03f957853;hb=900825c10932c82d66dcaaaabdedf3cb0f22928f;hpb=95771c01e72fca4414c3b4d75734126d7c53f6df 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 ec3f375bde..1c30fe2317 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 @@ -17,14 +17,23 @@ 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 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; @@ -33,12 +42,15 @@ 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; @@ -87,13 +99,15 @@ 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 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 @@ -101,17 +115,16 @@ 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 final Procedure createSnapshotProcedure = new CreateSnapshotProcedure(); private Stopwatch recoveryTimer; private int currentRecoveryBatchCount; + private final BehaviorStateHolder reusableBehaviorStateHolder = new BehaviorStateHolder(); + public RaftActor(String id, Map peerAddresses) { this(id, peerAddresses, Optional.absent()); } @@ -120,10 +133,11 @@ 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); + + context.setReplicatedLog(ReplicatedLogImpl.newInstance(context, delegatingPersistenceProvider, currentBehavior)); } private void initRecoveryTimer() { @@ -166,7 +180,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } else if (message instanceof ApplyJournalEntries) { onRecoveredApplyLogEntries(((ApplyJournalEntries) message).getToIndex()); } else if (message instanceof DeleteEntries) { - replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex()); + replicatedLog().removeFrom(((DeleteEntries) message).getFromIndex()); } else if (message instanceof UpdateElectionTerm) { context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor()); @@ -201,9 +215,9 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { // 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.setReplicatedLog(ReplicatedLogImpl.newInstance(snapshot, context, delegatingPersistenceProvider, + currentBehavior)); context.setLastApplied(snapshot.getLastAppliedIndex()); context.setCommitIndex(snapshot.getLastAppliedIndex()); @@ -214,8 +228,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { timer.stop(); LOG.info("Recovery snapshot applied for {} in {}: snapshotIndex={}, snapshotTerm={}, journal-size=" + - replicatedLog.size(), persistenceId(), timer.toString(), - replicatedLog.getSnapshotIndex(), replicatedLog.getSnapshotTerm()); + replicatedLog().size(), persistenceId(), timer.toString(), + replicatedLog().getSnapshotIndex(), replicatedLog().getSnapshotTerm()); } private void onRecoveredJournalLogEntry(ReplicatedLogEntry logEntry) { @@ -223,7 +237,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { LOG.debug("{}: Received ReplicatedLogEntry for recovery: {}", persistenceId(), logEntry.getIndex()); } - replicatedLog.append(logEntry); + replicatedLog().append(logEntry); } private void onRecoveredApplyLogEntries(long toIndex) { @@ -233,7 +247,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } for (long i = context.getLastApplied() + 1; i <= toIndex; i++) { - batchRecoveredLogEntry(replicatedLog.get(i)); + batchRecoveredLogEntry(replicatedLog().get(i)); } context.setLastApplied(toIndex); @@ -279,8 +293,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { "Persistence Id = " + persistenceId() + " Last index in log={}, snapshotIndex={}, snapshotTerm={}, " + "journal-size={}", - replicatedLog.lastIndex(), replicatedLog.getSnapshotIndex(), - replicatedLog.getSnapshotTerm(), replicatedLog.size()); + replicatedLog().lastIndex(), replicatedLog().getSnapshotIndex(), + replicatedLog().getSnapshotTerm(), replicatedLog().size()); initializeBehavior(); } @@ -290,9 +304,9 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected void changeCurrentBehavior(RaftActorBehavior newBehavior){ - RaftActorBehavior oldBehavior = currentBehavior; - currentBehavior = newBehavior; - handleBehaviorChange(oldBehavior, currentBehavior); + reusableBehaviorStateHolder.init(getCurrentBehavior()); + setCurrentBehavior(newBehavior); + handleBehaviorChange(reusableBehaviorStateHolder, getCurrentBehavior()); } @Override public void handleCommand(Object message) { @@ -319,11 +333,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { if(LOG.isDebugEnabled()) { LOG.debug("{}: Persisting ApplyLogEntries with index={}", persistenceId(), applyEntries.getToIndex()); } - persistence().persist(applyEntries, new Procedure() { - @Override - public void apply(ApplyJournalEntries param) throws Exception { - } - }); + + persistence().persist(applyEntries, NoopProcedure.instance()); } else if(message instanceof ApplySnapshot ) { Snapshot snapshot = ((ApplySnapshot) message).getSnapshot(); @@ -338,8 +349,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { applySnapshot(snapshot.getState()); //clears the followers log, sets the snapshot index to ensure adjusted-index works - replicatedLog = new ReplicatedLogImpl(snapshot); - context.setReplicatedLog(replicatedLog); + context.setReplicatedLog(ReplicatedLogImpl.newInstance(snapshot, context, delegatingPersistenceProvider, + currentBehavior)); context.setLastApplied(snapshot.getLastAppliedIndex()); } else if (message instanceof FindLeader) { @@ -362,49 +373,95 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { 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()); + context.getSnapshotManager().rollback(); } else if (message instanceof CaptureSnapshot) { - LOG.info("{}: CaptureSnapshot received by actor", persistenceId()); + LOG.debug("{}: CaptureSnapshot received by actor: {}", persistenceId(), message); - if(captureSnapshot == null) { - captureSnapshot = (CaptureSnapshot)message; - createSnapshot(); - } + context.getSnapshotManager().create(createSnapshotProcedure); - } else if (message instanceof CaptureSnapshotReply){ + } else if (message instanceof CaptureSnapshotReply) { handleCaptureSnapshotReply(((CaptureSnapshotReply) message).getSnapshot()); - + } else if(message instanceof GetOnDemandRaftState) { + onGetOnDemandRaftStats(); + } else if (message.equals(COMMIT_SNAPSHOT)) { + commitSnapshot(-1); } else { - RaftActorBehavior oldBehavior = currentBehavior; - currentBehavior = currentBehavior.handleMessage(getSender(), message); + 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()))); + } - handleBehaviorChange(oldBehavior, currentBehavior); + builder.followerInfoList(followerInfoList); } + + sender().tell(builder.build(), self()); + } - 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()); + } - if (getRoleChangeNotifier().isPresent() && + onLeaderChanged(oldBehaviorLeaderId, currentBehavior.getLeaderId()); + } + + 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()); } } @@ -429,46 +486,49 @@ 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 @@ -509,7 +569,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected ReplicatedLogEntry getLastLogEntry() { - return replicatedLog.last(); + return replicatedLog().last(); } protected Long getCurrentTerm(){ @@ -532,6 +592,41 @@ 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(COMMIT_SNAPSHOT, self()); + } + }); + } + } + /** * setPeerAddress sets the address of a known peer at a later time. *

@@ -551,10 +646,7 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } protected void commitSnapshot(long sequenceNumber) { - context.getReplicatedLog().snapshotCommit(); - - // TODO: Not sure if we want to be this aggressive with trimming stuff - trimPersistentData(sequenceNumber); + context.getSnapshotManager().commit(persistence(), sequenceNumber); } /** @@ -636,8 +728,6 @@ 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. @@ -646,17 +736,6 @@ 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(); @@ -675,202 +754,13 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } 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()); - - 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(), captureSnapshot.getLastAppliedIndex(), - captureSnapshot.getLastAppliedTerm()); + LOG.debug("{}: CaptureSnapshotReply received by actor: snapshot size {}", persistenceId(), snapshotBytes.length); - 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); + context.getSnapshotManager().persist(persistence(), snapshotBytes, currentBehavior, context.getTotalMemory()); } 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 { @@ -886,46 +776,6 @@ 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; @@ -945,38 +795,29 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { } } - protected class NonPersistentRaftDataProvider extends NonPersistentDataProvider { + private class CreateSnapshotProcedure implements Procedure { - public NonPersistentRaftDataProvider(){ - - } - - /** - * 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); + public void apply(Void aVoid) throws Exception { + createSnapshot(); } } - @VisibleForTesting - void setCurrentBehavior(AbstractRaftActorBehavior behavior) { - currentBehavior = behavior; - } + private static class BehaviorStateHolder { + private RaftActorBehavior behavior; + private String leaderId; - protected RaftActorBehavior getCurrentBehavior() { - return currentBehavior; - } + void init(RaftActorBehavior behavior) { + this.behavior = behavior; + this.leaderId = behavior != null ? behavior.getLeaderId() : null; + } + RaftActorBehavior getBehavior() { + return behavior; + } + + String getLeaderId() { + return leaderId; + } + } }