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=a7c3db4fc246bcce77d23e8d324a4301d930568e;hp=042b9fb56995d259d99a1559d4a3e3d5389fdba3;hb=887b1ccc82702af44d2bba4f3b45881410eaacf5;hpb=1b69a9ae877e79ba6addb1b0e343692b2acff1ec 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 042b9fb569..a7c3db4fc2 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 @@ -22,8 +22,11 @@ import com.google.common.annotations.VisibleForTesting; import com.google.common.base.Optional; import com.google.common.base.Stopwatch; import com.google.protobuf.ByteString; +import java.io.Serializable; +import java.util.Map; import org.opendaylight.controller.cluster.DataPersistenceProvider; import org.opendaylight.controller.cluster.common.actor.AbstractUntypedPersistentActor; +import org.opendaylight.controller.cluster.notifications.RoleChanged; import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries; import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot; import org.opendaylight.controller.cluster.raft.base.messages.ApplyState; @@ -41,9 +44,6 @@ import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply; import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload; import org.opendaylight.controller.protobuff.messages.cluster.raft.AppendEntriesMessages; -import java.io.Serializable; -import java.util.Map; - /** * RaftActor encapsulates a state machine that needs to be kept synchronized * in a cluster. It implements the RAFT algorithm as described in the paper @@ -113,6 +113,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { private int currentRecoveryBatchCount; + + public RaftActor(String id, Map peerAddresses) { this(id, peerAddresses, Optional.absent()); } @@ -169,8 +171,10 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { deleteSnapshots(new SnapshotSelectionCriteria(scala.Long.MaxValue(), scala.Long.MaxValue())); onRecoveryComplete(); + + RaftActorBehavior oldBehavior = currentBehavior; currentBehavior = new Follower(context); - onStateChanged(); + handleBehaviorChange(oldBehavior, currentBehavior); } } } @@ -269,8 +273,9 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { replicatedLog.lastIndex(), replicatedLog.snapshotIndex, replicatedLog.snapshotTerm, replicatedLog.size()); + RaftActorBehavior oldBehavior = currentBehavior; currentBehavior = new Follower(context); - onStateChanged(); + handleBehaviorChange(oldBehavior, currentBehavior); } @Override public void handleCommand(Object message) { @@ -366,26 +371,26 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { RaftActorBehavior oldBehavior = currentBehavior; currentBehavior = currentBehavior.handleMessage(getSender(), message); - if(oldBehavior != currentBehavior){ - onStateChanged(); - } - - onLeaderChanged(oldBehavior.getLeaderId(), currentBehavior.getLeaderId()); + handleBehaviorChange(oldBehavior, currentBehavior); } } - public java.util.Set getPeers() { - - return context.getPeerAddresses().keySet(); - } + private void handleBehaviorChange(RaftActorBehavior oldBehavior, RaftActorBehavior currentBehavior) { + if (oldBehavior != currentBehavior){ + onStateChanged(); + } + if (oldBehavior != null) { + // it can happen that the state has not changed but the leader has changed. + onLeaderChanged(oldBehavior.getLeaderId(), currentBehavior.getLeaderId()); - protected String getReplicatedLogState() { - return "snapshotIndex=" + context.getReplicatedLog().getSnapshotIndex() - + ", snapshotTerm=" + context.getReplicatedLog().getSnapshotTerm() - + ", im-mem journal size=" + context.getReplicatedLog().size(); + if (getRoleChangeNotifier().isPresent() && oldBehavior.state() != currentBehavior.state()) { + // we do not want to notify when the behavior/role is set for the first time (i.e follower) + getRoleChangeNotifier().get().tell(new RoleChanged(getId(), oldBehavior.state().name(), + currentBehavior.state().name()), getSelf()); + } + } } - /** * When a derived RaftActor needs to persist something it must call * persistData. @@ -394,8 +399,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { * @param identifier * @param data */ - protected void persistData(ActorRef clientActor, String identifier, - Payload data) { + protected void persistData(final ActorRef clientActor, final String identifier, + final Payload data) { ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry( context.getReplicatedLog().lastIndex() + 1, @@ -405,9 +410,42 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { LOG.debug("Persist data {}", replicatedLogEntry); } + final RaftActorContext raftContext = getRaftActorContext(); + replicatedLog - .appendAndPersist(clientActor, identifier, replicatedLogEntry); - } + .appendAndPersist(replicatedLogEntry, new Procedure() { + @Override + public void apply(ReplicatedLogEntry replicatedLogEntry) throws Exception { + if(!hasFollowers()){ + // Increment the Commit Index and the Last Applied values + raftContext.setCommitIndex(replicatedLogEntry.getIndex()); + raftContext.setLastApplied(replicatedLogEntry.getIndex()); + + // Apply the state immediately + applyState(clientActor, identifier, data); + + // Send a ApplyLogEntries message so that we write the fact that we applied + // the state to durable storage + self().tell(new ApplyLogEntries((int) replicatedLogEntry.getIndex()), self()); + + // Check if the "real" snapshot capture has been initiated. If no then do the fake snapshot + if(!hasSnapshotCaptureInitiated){ + raftContext.getReplicatedLog().snapshotPreCommit(raftContext.getLastApplied(), + raftContext.getTermInformation().getCurrentTerm()); + raftContext.getReplicatedLog().snapshotCommit(); + } else { + LOG.debug("Skipping fake snapshotting for {} because real snapshotting is in progress", getId()); + } + } else if (clientActor != null) { + // Send message for replication + currentBehavior.handleMessage(getSelf(), + new Replicate(clientActor, identifier, + replicatedLogEntry) + ); + } + + } + }); } protected String getId() { return context.getId(); @@ -578,6 +616,12 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { protected abstract DataPersistenceProvider persistence(); + /** + * Notifier Actor for this RaftActor to notify when a role change happens + * @return ActorRef - ActorRef of the notifier or Optional.absent if none. + */ + protected abstract Optional getRoleChangeNotifier(); + protected void onLeaderChanged(String oldLeader, String newLeader){}; private void trimPersistentData(long sequenceNumber) { @@ -641,8 +685,15 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { hasSnapshotCaptureInitiated = 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()); @@ -667,18 +718,27 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { @Override public void apply(DeleteEntries param) throws Exception { //FIXME : Doing nothing for now + dataSize = 0; + for(ReplicatedLogEntry entry : journal){ + dataSize += entry.size(); + } } }); } @Override public void appendAndPersist( final ReplicatedLogEntry replicatedLogEntry) { - appendAndPersist(null, null, replicatedLogEntry); + appendAndPersist(replicatedLogEntry, null); } - public void appendAndPersist(final ActorRef clientActor, - final String identifier, - final ReplicatedLogEntry replicatedLogEntry) { + @Override + public int dataSize() { + return dataSize; + } + + public void appendAndPersist( + final ReplicatedLogEntry replicatedLogEntry, + final Procedure callback) { if(LOG.isDebugEnabled()) { LOG.debug("Append log entry and persist {} ", replicatedLogEntry); @@ -696,16 +756,48 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { 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 (hasSnapshotCaptureInitiated == false && - journal.size() % context.getConfigParams().getSnapshotBatchCount() == 0) { + ( journalSize % context.getConfigParams().getSnapshotBatchCount() == 0 || + dataSizeForCheck > dataThreshold)) { + + dataSizeSinceLastSnapshot = 0; LOG.info("Initiating Snapshot Capture.."); long lastAppliedIndex = -1; long lastAppliedTerm = -1; ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied()); - if (lastAppliedEntry != null) { + if (!hasFollowers()) { + lastAppliedIndex = replicatedLogEntry.getIndex(); + lastAppliedTerm = replicatedLogEntry.getTerm(); + } else if (lastAppliedEntry != null) { lastAppliedIndex = lastAppliedEntry.getIndex(); lastAppliedTerm = lastAppliedEntry.getTerm(); } @@ -724,12 +816,8 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { null); hasSnapshotCaptureInitiated = true; } - // Send message for replication - if (clientActor != null) { - currentBehavior.handleMessage(getSelf(), - new Replicate(clientActor, identifier, - replicatedLogEntry) - ); + if(callback != null){ + callback.apply(replicatedLogEntry); } } } @@ -843,4 +931,5 @@ public abstract class RaftActor extends AbstractUntypedPersistentActor { protected RaftActorBehavior getCurrentBehavior() { return currentBehavior; } + }