Bump upstream SNAPSHOTS
[controller.git] / opendaylight / md-sal / sal-akka-raft / src / main / java / org / opendaylight / controller / cluster / raft / behaviors / Candidate.java
index b36c41abf262b8c24d05f354091f7c4ec92a0827..569f6b3d245fe5a6c1021f0537f08b3e1ec36f6e 100644 (file)
@@ -5,24 +5,29 @@
  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
  * and is available at http://www.eclipse.org/legal/epl-v10.html
  */
-
 package org.opendaylight.controller.cluster.raft.behaviors;
 
 import akka.actor.ActorRef;
 import akka.actor.ActorSelection;
-import java.util.Set;
+import java.util.ArrayList;
+import java.util.Collection;
+import org.opendaylight.controller.cluster.raft.PeerInfo;
 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.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;
 
 /**
- * The behavior of a RaftActor when it is in the CandidateState
- * <p/>
+ * The behavior of a RaftActor when it is in the Candidate raft state.
+ *
+ * <p>
  * Candidates (§5.2):
  * <ul>
  * <li> On conversion to candidate, start election:
@@ -38,88 +43,104 @@ import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
  * <li> If election timeout elapses: start new election
  * </ul>
  */
-public class Candidate extends AbstractRaftActorBehavior {
+public final class Candidate extends AbstractRaftActorBehavior {
 
     private int voteCount;
 
     private final int votesRequired;
 
-    private final Set<String> peers;
+    private final Collection<String> votingPeers = new ArrayList<>();
 
-    public Candidate(RaftActorContext context) {
+    public Candidate(final RaftActorContext context) {
         super(context, RaftState.Candidate);
 
-        peers = context.getPeerAddresses().keySet();
-
-        if(LOG.isDebugEnabled()) {
-            LOG.debug("{}: Election: Candidate has following peers: {}", logName(), peers);
+        for (PeerInfo peer: context.getPeers()) {
+            if (peer.isVoting()) {
+                votingPeers.add(peer.getId());
+            }
         }
 
-        votesRequired = getMajorityVoteCount(peers.size());
+        log.debug("{}: Election: Candidate has following voting peers: {}", logName(), votingPeers);
+
+        votesRequired = getMajorityVoteCount(votingPeers.size());
 
         startNewTerm();
-        scheduleElection(electionDuration());
+
+        if (votingPeers.isEmpty()) {
+            actor().tell(ElectionTimeout.INSTANCE, actor());
+        } else {
+            scheduleElection(electionDuration());
+        }
     }
 
-    @Override protected RaftActorBehavior handleAppendEntries(ActorRef sender,
-        AppendEntries appendEntries) {
+    @Override
+    public String getLeaderId() {
+        return null;
+    }
+
+    @Override
+    public short getLeaderPayloadVersion() {
+        return -1;
+    }
+
+    @Override
+    protected RaftActorBehavior handleAppendEntries(final ActorRef sender, final AppendEntries appendEntries) {
+
+        log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
 
-        if(LOG.isDebugEnabled()) {
-            LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
+        // Some other candidate for the same term became a leader and sent us an append entry
+        if (currentTerm() == appendEntries.getTerm()) {
+            log.info("{}: New Leader {} sent an AppendEntries to Candidate for term {} - will switch to Follower",
+                    logName(), appendEntries.getLeaderId(), currentTerm());
+
+            return switchBehavior(new Follower(context));
         }
 
         return this;
     }
 
-    @Override protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
-        AppendEntriesReply appendEntriesReply) {
-
+    @Override
+    protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
+            final AppendEntriesReply appendEntriesReply) {
         return this;
     }
 
-    @Override protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
-            RequestVoteReply requestVoteReply) {
-
-        LOG.debug("{}: handleRequestVoteReply: {}, current voteCount: {}", logName(), requestVoteReply,
-                voteCount);
+    @Override
+    protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender, final RequestVoteReply requestVoteReply) {
+        log.debug("{}: handleRequestVoteReply: {}, current voteCount: {}", logName(), requestVoteReply, voteCount);
 
         if (requestVoteReply.isVoteGranted()) {
             voteCount++;
         }
 
         if (voteCount >= votesRequired) {
-            return switchBehavior(new Leader(context));
+            if (context.getLastApplied() < context.getReplicatedLog().lastIndex()) {
+                log.info("{}: LastApplied index {} is behind last index {} - switching to PreLeader",
+                        logName(), context.getLastApplied(), context.getReplicatedLog().lastIndex());
+                return internalSwitchBehavior(RaftState.PreLeader);
+            } else {
+                return internalSwitchBehavior(RaftState.Leader);
+            }
         }
 
         return this;
     }
 
     @Override
-    public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
-
-        Object message = fromSerializableMessage(originalMessage);
-
-        if (message instanceof RaftRPC) {
-
-            RaftRPC rpc = (RaftRPC) message;
+    protected FiniteDuration electionDuration() {
+        return super.electionDuration().$div(context.getConfigParams().getCandidateElectionTimeoutDivisor());
+    }
 
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: RaftRPC message received {}, my term is {}", logName(), rpc,
-                        context.getTermInformation().getCurrentTerm());
-            }
 
-            // 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 switchBehavior(new Follower(context));
-            }
-        }
+    @Override
+    ApplyState getApplyStateFor(final ReplicatedLogEntry entry) {
+        throw new IllegalStateException("A candidate should never attempt to apply " + entry);
+    }
 
+    @Override
+    public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
         if (message instanceof ElectionTimeout) {
-            LOG.debug("{}: Received ElectionTimeout", logName());
+            log.debug("{}: Received ElectionTimeout", logName());
 
             if (votesRequired == 0) {
                 // If there are no peers then we should be a Leader
@@ -128,20 +149,44 @@ public class Candidate extends AbstractRaftActorBehavior {
                 // who we do not know about (as a peer)
                 // to send a message to the candidate
 
-                return switchBehavior(new Leader(context));
+                return internalSwitchBehavior(RaftState.Leader);
             }
+
             startNewTerm();
             scheduleElection(electionDuration());
             return this;
         }
 
-        return super.handleMessage(sender, message);
-    }
+        if (message instanceof RaftRPC) {
 
+            RaftRPC rpc = (RaftRPC) message;
 
-    private void startNewTerm() {
+            log.debug("{}: RaftRPC message received {}, my term is {}", logName(), rpc,
+                        context.getTermInformation().getCurrentTerm());
 
+            // 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()) {
+                log.info("{}: Term {} in \"{}\" message is greater than Candidate's term {} - switching to Follower",
+                        logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
 
+                context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
+
+                // The raft paper does not say whether or not a Candidate can/should process a RequestVote in
+                // this case but doing so gains quicker convergence when the sender's log is more up-to-date.
+                if (message instanceof RequestVote) {
+                    super.handleMessage(sender, message);
+                }
+
+                return internalSwitchBehavior(RaftState.Follower);
+            }
+        }
+
+        return super.handleMessage(sender, message);
+    }
+
+    private void startNewTerm() {
         // set voteCount back to 1 (that is voting for self)
         voteCount = 1;
 
@@ -150,28 +195,29 @@ public class Candidate extends AbstractRaftActorBehavior {
         long newTerm = currentTerm + 1;
         context.getTermInformation().updateAndPersist(newTerm, context.getId());
 
-        LOG.debug("{}: Starting new term {}", logName(), newTerm);
+        log.info("{}: Starting new election term {}", logName(), newTerm);
 
         // Request for a vote
         // TODO: Retry request for vote if replies do not arrive in a reasonable
         // amount of time TBD
-        for (String peerId : peers) {
+        for (String peerId : votingPeers) {
             ActorSelection peerActor = context.getPeerActorSelection(peerId);
-            if(peerActor != null) {
+            if (peerActor != null) {
                 RequestVote requestVote = new RequestVote(
                         context.getTermInformation().getCurrentTerm(),
                         context.getId(),
                         context.getReplicatedLog().lastIndex(),
                         context.getReplicatedLog().lastTerm());
 
-                LOG.debug("{}: Sending {} to peer {}", logName(), requestVote, peerId);
+                log.debug("{}: Sending {} to peer {}", logName(), requestVote, peerId);
 
                 peerActor.tell(requestVote, context.getActor());
             }
         }
     }
 
-    @Override public void close() throws Exception {
+    @Override
+    public void close() {
         stopElection();
     }
 }