import akka.actor.ActorRef;
import akka.actor.Cancellable;
+import akka.event.LoggingAdapter;
+import java.util.Random;
+import java.util.concurrent.TimeUnit;
import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
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.ElectionTimeout;
+import org.opendaylight.controller.cluster.raft.SerializationUtils;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyLogEntries;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
+import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
-import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
import org.opendaylight.controller.cluster.raft.messages.RequestVote;
import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
import scala.concurrent.duration.FiniteDuration;
-import java.util.Random;
-import java.util.concurrent.TimeUnit;
-
/**
* Abstract class that represents the behavior of a RaftActor
* <p/>
protected final RaftActorContext context;
/**
- * The maximum election time variance
- */
- private static final int ELECTION_TIME_MAX_VARIANCE = 100;
-
- /**
- * The interval at which a heart beat message will be sent to the remote
- * RaftActor
- * <p/>
- * Since this is set to 100 milliseconds the Election timeout should be
- * at least 200 milliseconds
- */
- protected static final FiniteDuration HEART_BEAT_INTERVAL =
- new FiniteDuration(100, TimeUnit.MILLISECONDS);
-
- /**
- * The interval in which a new election would get triggered if no leader is found
+ *
*/
- private static final long ELECTION_TIME_INTERVAL =
- HEART_BEAT_INTERVAL.toMillis() * 2;
+ protected final LoggingAdapter LOG;
/**
*
protected AbstractRaftActorBehavior(RaftActorContext context) {
this.context = context;
+ this.LOG = context.getLogger();
}
/**
*
* @param sender The actor that sent this message
* @param appendEntries The AppendEntries message
- * @param suggestedState The state that the RaftActor should be in based
- * on the base class's processing of the AppendEntries
- * message
* @return
*/
- protected abstract RaftState handleAppendEntries(ActorRef sender,
- AppendEntries appendEntries, RaftState suggestedState);
+ protected abstract RaftActorBehavior handleAppendEntries(ActorRef sender,
+ AppendEntries appendEntries);
- protected RaftState appendEntries(ActorRef sender,
- AppendEntries appendEntries, RaftState raftState) {
-
- if (raftState != state()) {
- context.getLogger().debug("Suggested state is " + raftState
- + " current behavior state is " + state());
- }
+ /**
+ * appendEntries first processes the AppendEntries message and then
+ * delegates handling to a specific behavior
+ *
+ * @param sender
+ * @param appendEntries
+ * @return
+ */
+ protected RaftActorBehavior appendEntries(ActorRef sender,
+ AppendEntries appendEntries) {
// 1. Reply false if term < currentTerm (§5.1)
if (appendEntries.getTerm() < currentTerm()) {
- context.getLogger().debug(
- "Cannot append entries because sender term " + appendEntries
- .getTerm() + " is less than " + currentTerm());
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("Cannot append entries because sender term {} is less than {}",
+ appendEntries.getTerm(), currentTerm());
+ }
+
sender.tell(
new AppendEntriesReply(context.getId(), currentTerm(), false,
lastIndex(), lastTerm()), actor()
);
- return state();
+ return this;
}
- return handleAppendEntries(sender, appendEntries, raftState);
+ return handleAppendEntries(sender, appendEntries);
}
/**
*
* @param sender The actor that sent this message
* @param appendEntriesReply The AppendEntriesReply message
- * @param suggestedState The state that the RaftActor should be in based
- * on the base class's processing of the
- * AppendEntriesReply message
* @return
*/
+ protected abstract RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
+ AppendEntriesReply appendEntriesReply);
- protected abstract RaftState handleAppendEntriesReply(ActorRef sender,
- AppendEntriesReply appendEntriesReply, RaftState suggestedState);
+ /**
+ * requestVote handles the RequestVote message. This logic is common
+ * for all behaviors
+ *
+ * @param sender
+ * @param requestVote
+ * @return
+ */
+ protected RaftActorBehavior requestVote(ActorRef sender,
+ RequestVote requestVote) {
- protected RaftState requestVote(ActorRef sender,
- RequestVote requestVote, RaftState suggestedState) {
+ if(LOG.isDebugEnabled()) {
+ LOG.debug(requestVote.toString());
+ }
boolean grantVote = false;
if (candidateLatest) {
grantVote = true;
- context.getTermInformation().update(requestVote.getTerm(),
+ context.getTermInformation().updateAndPersist(requestVote.getTerm(),
requestVote.getCandidateId());
}
}
sender.tell(new RequestVoteReply(currentTerm(), grantVote), actor());
- return suggestedState;
+ return this;
}
/**
*
* @param sender The actor that sent this message
* @param requestVoteReply The RequestVoteReply message
- * @param suggestedState The state that the RaftActor should be in based
- * on the base class's processing of the RequestVote
- * message
* @return
*/
+ protected abstract RaftActorBehavior handleRequestVoteReply(ActorRef sender,
+ RequestVoteReply requestVoteReply);
- protected abstract RaftState handleRequestVoteReply(ActorRef sender,
- RequestVoteReply requestVoteReply, RaftState suggestedState);
-
+ /**
+ * Creates a random election duration
+ *
+ * @return
+ */
protected FiniteDuration electionDuration() {
- long variance = new Random().nextInt(ELECTION_TIME_MAX_VARIANCE);
- return new FiniteDuration(ELECTION_TIME_INTERVAL + variance,
- TimeUnit.MILLISECONDS);
+ long variance = new Random().nextInt(context.getConfigParams().getElectionTimeVariance());
+ return context.getConfigParams().getElectionTimeOutInterval().$plus(
+ new FiniteDuration(variance, TimeUnit.MILLISECONDS));
}
+ /**
+ * stop the scheduled election
+ */
protected void stopElection() {
if (electionCancel != null && !electionCancel.isCancelled()) {
electionCancel.cancel();
}
}
+ /**
+ * schedule a new election
+ *
+ * @param interval
+ */
protected void scheduleElection(FiniteDuration interval) {
-
stopElection();
// Schedule an election. When the scheduler triggers an ElectionTimeout
context.getActorSystem().dispatcher(), context.getActor());
}
+ /**
+ * Get the current term
+ * @return
+ */
protected long currentTerm() {
return context.getTermInformation().getCurrentTerm();
}
+ /**
+ * Get the candidate for whom we voted in the current term
+ * @return
+ */
protected String votedFor() {
return context.getTermInformation().getVotedFor();
}
+ /**
+ * Get the actor associated with this behavior
+ * @return
+ */
protected ActorRef actor() {
return context.getActor();
}
+ /**
+ * Get the term from the last entry in the log
+ *
+ * @return
+ */
protected long lastTerm() {
return context.getReplicatedLog().lastTerm();
}
+ /**
+ * Get the index from the last entry in the log
+ *
+ * @return
+ */
protected long lastIndex() {
return context.getReplicatedLog().lastIndex();
}
+ /**
+ * Find the client request tracker for a specific logIndex
+ *
+ * @param logIndex
+ * @return
+ */
protected ClientRequestTracker findClientRequestTracker(long logIndex) {
return null;
}
- protected void applyLogToStateMachine(long index) {
+ /**
+ * Find the client request tracker for a specific logIndex
+ *
+ * @param logIndex
+ * @return
+ */
+ protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
+ return null;
+ }
+
+
+ /**
+ * Find the log index from the previous to last entry in the log
+ *
+ * @return
+ */
+ protected long prevLogIndex(long index){
+ ReplicatedLogEntry prevEntry =
+ context.getReplicatedLog().get(index - 1);
+ if (prevEntry != null) {
+ return prevEntry.getIndex();
+ }
+ return -1;
+ }
+
+ /**
+ * Find the log term from the previous to last entry in the log
+ * @return
+ */
+ protected long prevLogTerm(long index){
+ ReplicatedLogEntry prevEntry =
+ context.getReplicatedLog().get(index - 1);
+ if (prevEntry != null) {
+ return prevEntry.getTerm();
+ }
+ return -1;
+ }
+
+ /**
+ * Apply the provided index to the state machine
+ *
+ * @param index a log index that is known to be committed
+ */
+ protected void applyLogToStateMachine(final long index) {
+ long newLastApplied = context.getLastApplied();
// Now maybe we apply to the state machine
for (long i = context.getLastApplied() + 1;
i < index + 1; i++) {
ActorRef clientActor = null;
String identifier = null;
- ClientRequestTracker tracker = findClientRequestTracker(i);
+ ClientRequestTracker tracker = removeClientRequestTracker(i);
if (tracker != null) {
clientActor = tracker.getClientActor();
context.getReplicatedLog().get(i);
if (replicatedLogEntry != null) {
+ // Send a local message to the local RaftActor (it's derived class to be
+ // specific to apply the log to it's index)
actor().tell(new ApplyState(clientActor, identifier,
replicatedLogEntry), actor());
+ newLastApplied = i;
} else {
- context.getLogger().error(
- "Missing index " + i + " from log. Cannot apply state.");
+ //if one index is not present in the log, no point in looping
+ // around as the rest wont be present either
+ LOG.warning(
+ "Missing index {} from log. Cannot apply state. Ignoring {} to {}", i, i, index);
+ break;
}
}
- // Send a local message to the local RaftActor (it's derived class to be
- // specific to apply the log to it's index)
- context.setLastApplied(index);
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("Setting last applied to {}", newLastApplied);
+ }
+ context.setLastApplied(newLastApplied);
+
+ // send a message to persist a ApplyLogEntries marker message into akka's persistent journal
+ // will be used during recovery
+ //in case if the above code throws an error and this message is not sent, it would be fine
+ // as the append entries received later would initiate add this message to the journal
+ actor().tell(new ApplyLogEntries((int) context.getLastApplied()), actor());
+ }
+
+ protected Object fromSerializableMessage(Object serializable){
+ return SerializationUtils.fromSerializable(serializable);
}
@Override
- public RaftState handleMessage(ActorRef sender, Object message) {
- RaftState raftState = state();
- if (message instanceof RaftRPC) {
- raftState = applyTerm((RaftRPC) message);
- }
+ public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
if (message instanceof AppendEntries) {
- raftState = appendEntries(sender, (AppendEntries) message,
- raftState);
+ return appendEntries(sender, (AppendEntries) message);
} else if (message instanceof AppendEntriesReply) {
- raftState =
- handleAppendEntriesReply(sender, (AppendEntriesReply) message,
- raftState);
+ return handleAppendEntriesReply(sender, (AppendEntriesReply) message);
} else if (message instanceof RequestVote) {
- raftState =
- requestVote(sender, (RequestVote) message, raftState);
+ return requestVote(sender, (RequestVote) message);
} else if (message instanceof RequestVoteReply) {
- raftState =
- handleRequestVoteReply(sender, (RequestVoteReply) message,
- raftState);
+ return handleRequestVoteReply(sender, (RequestVoteReply) message);
}
- return raftState;
+ return this;
}
@Override public String getLeaderId() {
return leaderId;
}
- private RaftState applyTerm(RaftRPC rpc) {
- // 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().update(rpc.getTerm(), null);
- return RaftState.Follower;
+ protected RaftActorBehavior switchBehavior(RaftActorBehavior behavior) {
+ LOG.info("{} :- Switching from behavior {} to {}", context.getId(), this.state(), behavior.state());
+ try {
+ close();
+ } catch (Exception e) {
+ LOG.error(e, "Failed to close behavior : {}", this.state());
}
- return state();
+
+ return behavior;
}
+ protected int getMajorityVoteCount(int numPeers) {
+ // Votes are required from a majority of the peers including self.
+ // The numMajority field therefore stores a calculated value
+ // of the number of votes required for this candidate to win an
+ // election based on it's known peers.
+ // If a peer was added during normal operation and raft replicas
+ // came to know about them then the new peer would also need to be
+ // taken into consideration when calculating this value.
+ // Here are some examples for what the numMajority would be for n
+ // peers
+ // 0 peers = 1 numMajority -: (0 + 1) / 2 + 1 = 1
+ // 2 peers = 2 numMajority -: (2 + 1) / 2 + 1 = 2
+ // 4 peers = 3 numMajority -: (4 + 1) / 2 + 1 = 3
+
+ int numMajority = 0;
+ if (numPeers > 0) {
+ int self = 1;
+ numMajority = (numPeers + self) / 2 + 1;
+ }
+ return numMajority;
+
+ }
}