package org.opendaylight.controller.cluster.raft.behaviors;
import akka.actor.ActorRef;
+import akka.actor.ActorSelection;
+import akka.actor.Address;
+import akka.cluster.Cluster;
+import akka.cluster.ClusterEvent.CurrentClusterState;
+import akka.cluster.Member;
+import akka.cluster.MemberStatus;
import akka.japi.Procedure;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.Stopwatch;
import java.util.ArrayList;
+import java.util.Optional;
+import java.util.Set;
import java.util.concurrent.TimeUnit;
import javax.annotation.Nullable;
import org.opendaylight.controller.cluster.raft.RaftActorContext;
public class Follower extends AbstractRaftActorBehavior {
private static final int SYNC_THRESHOLD = 10;
+ private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
+
private final SyncStatusTracker initialSyncStatusTracker;
- private final Procedure<ReplicatedLogEntry> appendAndPersistCallback = new Procedure<ReplicatedLogEntry>() {
- @Override
- public void apply(ReplicatedLogEntry logEntry) {
- context.getReplicatedLog().captureSnapshotIfReady(logEntry);
- }
- };
+ private final Procedure<ReplicatedLogEntry> appendAndPersistCallback =
+ logEntry -> context.getReplicatedLog().captureSnapshotIfReady(logEntry);
- private final Stopwatch lastLeaderMessageTimer = Stopwatch.createUnstarted();
+ private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted();
private SnapshotTracker snapshotTracker = null;
private String leaderId;
private short leaderPayloadVersion;
// to make it easier to read. Before refactoring ensure tests
// cover the code properly
+ if (snapshotTracker != null && !snapshotTracker.getLeaderId().equals(appendEntries.getLeaderId())) {
+ LOG.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the " +
+ "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId());
+ snapshotTracker = null;
+ }
+
if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
// if snapshot install is in progress, follower should just acknowledge append entries with a reply.
AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
}
@Override
- public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
- if (originalMessage instanceof ElectionTimeout || originalMessage instanceof TimeoutNow) {
- return handleElectionTimeout(originalMessage);
+ public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
+ if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
+ return handleElectionTimeout(message);
}
- final Object message = fromSerializableMessage(originalMessage);
if (!(message instanceof RaftRPC)) {
// The rest of the processing requires the message to be a RaftRPC
return null;
// queue but would be processed before the ElectionTimeout message and thus would restart the
// lastLeaderMessageTimer.
long lastLeaderMessageInterval = lastLeaderMessageTimer.elapsed(TimeUnit.MILLISECONDS);
- boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning() || lastLeaderMessageInterval >=
- context.getConfigParams().getElectionTimeOutInterval().toMillis();
+ long electionTimeoutInMillis = context.getConfigParams().getElectionTimeOutInterval().toMillis();
+ boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning() ||
+ lastLeaderMessageInterval >= electionTimeoutInMillis;
if(canStartElection()) {
- if(message instanceof TimeoutNow || noLeaderMessageReceived) {
- LOG.debug("{}: Received {} - switching to Candidate", logName(), message.getClass().getSimpleName());
+ if(message instanceof TimeoutNow) {
+ LOG.debug("{}: Received TimeoutNow - switching to Candidate", logName());
return internalSwitchBehavior(RaftState.Candidate);
+ } else if(noLeaderMessageReceived) {
+ // Check the cluster state to see if the leader is known to be up before we go to Candidate.
+ // However if we haven't heard from the leader in a long time even though the cluster state
+ // indicates it's up then something is wrong - leader might be stuck indefinitely - so switch
+ // to Candidate,
+ long maxElectionTimeout = electionTimeoutInMillis * MAX_ELECTION_TIMEOUT_FACTOR;
+ if(isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
+ LOG.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
+ scheduleElection(electionDuration());
+ } else {
+ LOG.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
+ return internalSwitchBehavior(RaftState.Candidate);
+ }
} else {
- LOG.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout",
- logName(), lastLeaderMessageInterval);
+ LOG.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
+ logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
scheduleElection(electionDuration());
}
} else if(message instanceof ElectionTimeout) {
return this;
}
+ private boolean isLeaderAvailabilityKnown() {
+ if(leaderId == null) {
+ return false;
+ }
+
+ Optional<Cluster> cluster = context.getCluster();
+ if(!cluster.isPresent()) {
+ return false;
+ }
+
+ ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
+ if(leaderActor == null) {
+ return false;
+ }
+
+ Address leaderAddress = leaderActor.anchorPath().address();
+
+ CurrentClusterState state = cluster.get().state();
+ Set<Member> unreachable = state.getUnreachable();
+
+ LOG.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
+ unreachable);
+
+ for(Member m: unreachable) {
+ if(leaderAddress.equals(m.address())) {
+ LOG.info("{}: Leader {} is unreachable", logName(), leaderAddress);
+ return false;
+ }
+ }
+
+ for(Member m: state.getMembers()) {
+ if(leaderAddress.equals(m.address())) {
+ if(m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
+ LOG.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
+ leaderAddress, m.status());
+ return true;
+ } else {
+ LOG.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
+ leaderAddress, m.status());
+ return false;
+ }
+ }
+ }
+
+ LOG.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
+
+ return false;
+ }
+
private void handleInstallSnapshot(final ActorRef sender, InstallSnapshot installSnapshot) {
LOG.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
leaderId = installSnapshot.getLeaderId();
if(snapshotTracker == null){
- snapshotTracker = new SnapshotTracker(LOG, installSnapshot.getTotalChunks());
+ snapshotTracker = new SnapshotTracker(LOG, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId());
}
updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());