import akka.actor.ActorRef;
import akka.actor.Cancellable;
+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.SerializationUtils;
-import org.opendaylight.controller.cluster.raft.internal.messages.ApplyState;
-import org.opendaylight.controller.cluster.raft.internal.messages.ElectionTimeout;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
+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.RequestVote;
import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
+import org.slf4j.Logger;
import scala.concurrent.duration.FiniteDuration;
-import java.util.Random;
-import java.util.concurrent.TimeUnit;
-
/**
* Abstract class that represents the behavior of a RaftActor
* <p/>
*/
public abstract class AbstractRaftActorBehavior implements RaftActorBehavior {
+ protected static final ElectionTimeout ELECTION_TIMEOUT = new ElectionTimeout();
+
/**
* Information about the RaftActor whose behavior this class represents
*/
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 Logger LOG;
/**
*
*/
protected String leaderId = null;
+ private long replicatedToAllIndex = -1;
+
+ private final String logName;
- protected AbstractRaftActorBehavior(RaftActorContext context) {
+ private final RaftState state;
+
+ protected AbstractRaftActorBehavior(RaftActorContext context, RaftState state) {
this.context = context;
+ this.state = state;
+ this.LOG = context.getLogger();
+
+ logName = String.format("%s (%s)", context.getId(), state);
+ }
+
+ @Override
+ public RaftState state() {
+ return state;
+ }
+
+ public String logName() {
+ return logName;
+ }
+
+ @Override
+ public void setReplicatedToAllIndex(long replicatedToAllIndex) {
+ this.replicatedToAllIndex = replicatedToAllIndex;
+ }
+
+ @Override
+ public long getReplicatedToAllIndex() {
+ return replicatedToAllIndex;
}
/**
* @param appendEntries The AppendEntries message
* @return
*/
- protected abstract RaftState handleAppendEntries(ActorRef sender,
+ protected abstract RaftActorBehavior handleAppendEntries(ActorRef sender,
AppendEntries appendEntries);
* @param appendEntries
* @return
*/
- protected RaftState appendEntries(ActorRef sender,
+ protected RaftActorBehavior appendEntries(ActorRef sender,
AppendEntries appendEntries) {
// 1. Reply false if term < currentTerm (ยง5.1)
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 {}",
+ logName(), appendEntries.getTerm(), currentTerm());
+ }
+
sender.tell(
new AppendEntriesReply(context.getId(), currentTerm(), false,
lastIndex(), lastTerm()), actor()
);
- return state();
+ return this;
}
* @param appendEntriesReply The AppendEntriesReply message
* @return
*/
- protected abstract RaftState handleAppendEntriesReply(ActorRef sender,
+ protected abstract RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
AppendEntriesReply appendEntriesReply);
/**
* @param requestVote
* @return
*/
- protected RaftState requestVote(ActorRef sender,
- RequestVote requestVote) {
+ protected RaftActorBehavior requestVote(ActorRef sender, RequestVote requestVote) {
+
+ LOG.debug("{}: In requestVote: {}", logName(), requestVote);
boolean grantVote = false;
}
}
- sender.tell(new RequestVoteReply(currentTerm(), grantVote), actor());
+ RequestVoteReply reply = new RequestVoteReply(currentTerm(), grantVote);
+
+ LOG.debug("{}: requestVote returning: {}", logName(), reply);
+
+ sender.tell(reply, actor());
- return state();
+ return this;
}
/**
* @param requestVoteReply The RequestVoteReply message
* @return
*/
- protected abstract RaftState handleRequestVoteReply(ActorRef sender,
+ protected abstract RaftActorBehavior handleRequestVoteReply(ActorRef sender,
RequestVoteReply requestVoteReply);
/**
* @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));
}
/**
// message is sent to itself
electionCancel =
context.getActorSystem().scheduler().scheduleOnce(interval,
- context.getActor(), new ElectionTimeout(),
+ context.getActor(), ELECTION_TIMEOUT,
context.getActorSystem().dispatcher(), context.getActor());
}
return null;
}
+ /**
+ * 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
*
*
* @param index a log index that is known to be committed
*/
- protected void applyLogToStateMachine(long index) {
+ 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.warn(
+ "{}: Missing index {} from log. Cannot apply state. Ignoring {} to {}",
+ logName(), 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 {}", logName(), 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 ApplyJournalEntries(context.getLastApplied()), actor());
}
protected Object fromSerializableMessage(Object serializable){
}
@Override
- public RaftState handleMessage(ActorRef sender, Object message) {
+ public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
if (message instanceof AppendEntries) {
return appendEntries(sender, (AppendEntries) message);
} else if (message instanceof AppendEntriesReply) {
} else if (message instanceof RequestVoteReply) {
return handleRequestVoteReply(sender, (RequestVoteReply) message);
}
- return state();
+ return this;
}
@Override public String getLeaderId() {
return leaderId;
}
+
+ protected RaftActorBehavior switchBehavior(RaftActorBehavior behavior) {
+ LOG.info("{} :- Switching from behavior {} to {}", logName(), this.state(), behavior.state());
+ try {
+ close();
+ } catch (Exception e) {
+ LOG.error("{}: Failed to close behavior : {}", logName(), this.state(), e);
+ }
+
+ 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;
+
+ }
+
+
+ /**
+ * Performs a snapshot with no capture on the replicated log.
+ * It clears the log from the supplied index or last-applied-1 which ever is minimum.
+ *
+ * @param snapshotCapturedIndex
+ */
+ protected void performSnapshotWithoutCapture(final long snapshotCapturedIndex) {
+ // we would want to keep the lastApplied as its used while capturing snapshots
+ long lastApplied = context.getLastApplied();
+ long tempMin = Math.min(snapshotCapturedIndex, (lastApplied > -1 ? lastApplied - 1 : -1));
+
+ if(LOG.isTraceEnabled()) {
+ LOG.trace("{}: performSnapshotWithoutCapture: snapshotCapturedIndex: {}, lastApplied: {}, tempMin: {}",
+ logName, snapshotCapturedIndex, lastApplied, tempMin);
+ }
+
+ if (tempMin > -1 && context.getReplicatedLog().isPresent(tempMin)) {
+ LOG.debug("{}: fakeSnapshot purging log to {} for term {}", logName(), tempMin,
+ context.getTermInformation().getCurrentTerm());
+
+ //use the term of the temp-min, since we check for isPresent, entry will not be null
+ ReplicatedLogEntry entry = context.getReplicatedLog().get(tempMin);
+ context.getReplicatedLog().snapshotPreCommit(tempMin, entry.getTerm());
+ context.getReplicatedLog().snapshotCommit();
+ setReplicatedToAllIndex(tempMin);
+ } else if(tempMin > getReplicatedToAllIndex()) {
+ // It's possible a follower was lagging and an install snapshot advanced its match index past
+ // the current replicatedToAllIndex. Since the follower is now caught up we should advance the
+ // replicatedToAllIndex (to tempMin). The fact that tempMin wasn't found in the log is likely
+ // due to a previous snapshot triggered by the memory threshold exceeded, in that case we
+ // trim the log to the last applied index even if previous entries weren't replicated to all followers.
+ setReplicatedToAllIndex(tempMin);
+ }
+ }
+
+ protected String getId(){
+ return context.getId();
+ }
+
}
Code Review / controller.git / blobdiff