import org.opendaylight.controller.cluster.raft.RaftActorContext;
import org.opendaylight.controller.cluster.raft.RaftState;
import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
-import org.opendaylight.controller.cluster.raft.internal.messages.ApplyState;
-import org.opendaylight.controller.cluster.raft.internal.messages.Replicate;
-import org.opendaylight.controller.cluster.raft.internal.messages.SendHeartBeat;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
+import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
+import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
+import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
+import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
+import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
+import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
import scala.concurrent.duration.FiniteDuration;
import java.util.ArrayList;
+import java.util.Collections;
import java.util.HashMap;
import java.util.List;
import java.util.Map;
+import java.util.Set;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicLong;
private final Map<String, FollowerLogInformation> followerToLog =
new HashMap();
- private final Map<String, ActorSelection> followerToActor = new HashMap<>();
+ private final Set<String> followers;
- private Cancellable heartbeatCancel = null;
+ private Cancellable heartbeatSchedule = null;
+ private Cancellable appendEntriesSchedule = null;
+ private Cancellable installSnapshotSchedule = null;
private List<ClientRequestTracker> trackerList = new ArrayList<>();
context.setCommitIndex(lastIndex());
}
- for (String followerId : context.getPeerAddresses().keySet()) {
+ followers = context.getPeerAddresses().keySet();
+
+ for (String followerId : followers) {
FollowerLogInformation followerLogInformation =
new FollowerLogInformationImpl(followerId,
new AtomicLong(lastIndex()),
new AtomicLong(-1));
- followerToActor.put(followerId,
- context.actorSelection(context.getPeerAddress(followerId)));
-
followerToLog.put(followerId, followerLogInformation);
-
}
- if (followerToActor.size() > 0) {
- minReplicationCount = (followerToActor.size() + 1) / 2 + 1;
+ context.getLogger().debug("Election:Leader has following peers:"+ followers);
+
+ if (followers.size() > 0) {
+ minReplicationCount = (followers.size() + 1) / 2 + 1;
} else {
minReplicationCount = 0;
}
// prevent election timeouts (§5.2)
scheduleHeartBeat(new FiniteDuration(0, TimeUnit.SECONDS));
+ scheduleInstallSnapshotCheck(
+ new FiniteDuration(context.getConfigParams().getHeartBeatInterval().length() * 1000,
+ context.getConfigParams().getHeartBeatInterval().unit())
+ );
}
@Override protected RaftState handleAppendEntries(ActorRef sender,
- AppendEntries appendEntries, RaftState suggestedState) {
+ AppendEntries appendEntries) {
- context.getLogger()
- .error("An unexpected AppendEntries received in state " + state());
+ context.getLogger().info("Leader: Received {}", appendEntries.toString());
- return suggestedState;
+ return state();
}
@Override protected RaftState handleAppendEntriesReply(ActorRef sender,
- AppendEntriesReply appendEntriesReply, RaftState suggestedState) {
+ AppendEntriesReply appendEntriesReply) {
- // Do not take any other action since a behavior change is coming
- if (suggestedState != state())
- return suggestedState;
+ if(! appendEntriesReply.isSuccess()) {
+ context.getLogger()
+ .info("Leader: Received {}", appendEntriesReply.toString());
+ }
// Update the FollowerLogInformation
String followerId = appendEntriesReply.getFollowerId();
FollowerLogInformation followerLogInformation =
followerToLog.get(followerId);
+
+ if(followerLogInformation == null){
+ context.getLogger().error("Unknown follower {}", followerId);
+ return state();
+ }
+
if (appendEntriesReply.isSuccess()) {
followerLogInformation
.setMatchIndex(appendEntriesReply.getLogLastIndex());
followerLogInformation
.setNextIndex(appendEntriesReply.getLogLastIndex() + 1);
} else {
+
+ // TODO: When we find that the follower is out of sync with the
+ // Leader we simply decrement that followers next index by 1.
+ // Would it be possible to do better than this? The RAFT spec
+ // does not explicitly deal with it but may be something for us to
+ // think about
+
followerLogInformation.decrNextIndex();
}
applyLogToStateMachine(context.getCommitIndex());
}
- return suggestedState;
+ return state();
}
protected ClientRequestTracker findClientRequestTracker(long logIndex) {
}
@Override protected RaftState handleRequestVoteReply(ActorRef sender,
- RequestVoteReply requestVoteReply, RaftState suggestedState) {
- return suggestedState;
+ RequestVoteReply requestVoteReply) {
+ return state();
}
@Override public RaftState state() {
return RaftState.Leader;
}
- @Override public RaftState handleMessage(ActorRef sender, Object message) {
+ @Override public RaftState handleMessage(ActorRef sender, Object originalMessage) {
Preconditions.checkNotNull(sender, "sender should not be null");
+ Object message = fromSerializableMessage(originalMessage);
+
+ if (message instanceof RaftRPC) {
+ RaftRPC rpc = (RaftRPC) message;
+ // If RPC request or response contains term T > currentTerm:
+ // set currentTerm = T, convert to follower (§5.1)
+ // This applies to all RPC messages and responses
+ if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
+ context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
+ return RaftState.Follower;
+ }
+ }
+
try {
if (message instanceof SendHeartBeat) {
return sendHeartBeat();
+ } else if(message instanceof SendInstallSnapshot) {
+ installSnapshotIfNeeded();
} else if (message instanceof Replicate) {
replicate((Replicate) message);
+ } else if (message instanceof InstallSnapshotReply){
+ handleInstallSnapshotReply(
+ (InstallSnapshotReply) message);
}
} finally {
- scheduleHeartBeat(HEART_BEAT_INTERVAL);
+ scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
}
return super.handleMessage(sender, message);
}
+ private void handleInstallSnapshotReply(InstallSnapshotReply message) {
+ InstallSnapshotReply reply = message;
+ String followerId = reply.getFollowerId();
+ FollowerLogInformation followerLogInformation =
+ followerToLog.get(followerId);
+
+ followerLogInformation
+ .setMatchIndex(context.getReplicatedLog().getSnapshotIndex());
+ followerLogInformation
+ .setNextIndex(context.getReplicatedLog().getSnapshotIndex() + 1);
+ }
+
private void replicate(Replicate replicate) {
long logIndex = replicate.getReplicatedLogEntry().getIndex();
context.getLogger().debug("Replicate message " + logIndex);
- if (followerToActor.size() == 0) {
+ if (followers.size() == 0) {
context.setCommitIndex(
replicate.getReplicatedLogEntry().getIndex());
private void sendAppendEntries() {
// Send an AppendEntries to all followers
- for (String followerId : followerToActor.keySet()) {
+ for (String followerId : followers) {
ActorSelection followerActor =
- followerToActor.get(followerId);
+ context.getPeerActorSelection(followerId);
- FollowerLogInformation followerLogInformation =
- followerToLog.get(followerId);
+ if (followerActor != null) {
+ FollowerLogInformation followerLogInformation =
+ followerToLog.get(followerId);
- long nextIndex = followerLogInformation.getNextIndex().get();
+ long nextIndex = followerLogInformation.getNextIndex().get();
- List<ReplicatedLogEntry> entries =
- context.getReplicatedLog().getFrom(nextIndex);
+ List<ReplicatedLogEntry> entries = Collections.emptyList();
- followerActor.tell(
- new AppendEntries(currentTerm(), context.getId(),
- prevLogIndex(nextIndex), prevLogTerm(nextIndex),
- entries, context.getCommitIndex()
- ),
- actor()
- );
+ if (context.getReplicatedLog().isPresent(nextIndex)) {
+ // TODO: Instead of sending all entries from nextIndex
+ // only send a fixed number of entries to each follower
+ // This is to avoid the situation where there are a lot of
+ // entries to install for a fresh follower or to a follower
+ // that has fallen too far behind with the log but yet is not
+ // eligible to receive a snapshot
+ entries =
+ context.getReplicatedLog().getFrom(nextIndex);
+ }
+
+ followerActor.tell(
+ new AppendEntries(currentTerm(), context.getId(),
+ prevLogIndex(nextIndex),
+ prevLogTerm(nextIndex), entries,
+ context.getCommitIndex()).toSerializable(),
+ actor()
+ );
+ }
+ }
+ }
+
+ /**
+ * An installSnapshot is scheduled at a interval that is a multiple of
+ * a HEARTBEAT_INTERVAL. This is to avoid the need to check for installing
+ * snapshots at every heartbeat.
+ */
+ private void installSnapshotIfNeeded(){
+ for (String followerId : followers) {
+ ActorSelection followerActor =
+ context.getPeerActorSelection(followerId);
+
+ if(followerActor != null) {
+ FollowerLogInformation followerLogInformation =
+ followerToLog.get(followerId);
+
+ long nextIndex = followerLogInformation.getNextIndex().get();
+
+ if (!context.getReplicatedLog().isPresent(nextIndex) && context
+ .getReplicatedLog().isInSnapshot(nextIndex)) {
+ followerActor.tell(
+ new InstallSnapshot(currentTerm(), context.getId(),
+ context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().getSnapshotTerm(),
+ context.getReplicatedLog().getSnapshot()
+ ),
+ actor()
+ );
+ }
+ }
}
}
private RaftState sendHeartBeat() {
- if (followerToActor.size() > 0) {
+ if (followers.size() > 0) {
sendAppendEntries();
}
return state();
}
private void stopHeartBeat() {
- if (heartbeatCancel != null && !heartbeatCancel.isCancelled()) {
- heartbeatCancel.cancel();
+ if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
+ heartbeatSchedule.cancel();
+ }
+ }
+
+ private void stopInstallSnapshotSchedule() {
+ if (installSnapshotSchedule != null && !installSnapshotSchedule.isCancelled()) {
+ installSnapshotSchedule.cancel();
}
}
private void scheduleHeartBeat(FiniteDuration interval) {
- if(followerToActor.keySet().size() == 0){
+ if(followers.size() == 0){
// Optimization - do not bother scheduling a heartbeat as there are
// no followers
return;
// Scheduling the heartbeat only once here because heartbeats do not
// need to be sent if there are other messages being sent to the remote
// actor.
- heartbeatCancel =
+ heartbeatSchedule =
context.getActorSystem().scheduler().scheduleOnce(
interval,
context.getActor(), new SendHeartBeat(),
context.getActorSystem().dispatcher(), context.getActor());
}
+
+ private void scheduleInstallSnapshotCheck(FiniteDuration interval) {
+ if(followers.size() == 0){
+ // Optimization - do not bother scheduling a heartbeat as there are
+ // no followers
+ return;
+ }
+
+ stopInstallSnapshotSchedule();
+
+ // Schedule a message to send append entries to followers that can
+ // accept an append entries with some data in it
+ installSnapshotSchedule =
+ context.getActorSystem().scheduler().scheduleOnce(
+ interval,
+ context.getActor(), new SendInstallSnapshot(),
+ context.getActorSystem().dispatcher(), context.getActor());
+ }
+
+
+
@Override public void close() throws Exception {
stopHeartBeat();
}
Code Review / controller.git / blobdiff