* 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.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 edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
import java.io.IOException;
import java.util.ArrayList;
+import java.util.HashSet;
import java.util.List;
import java.util.Optional;
import java.util.Set;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicBoolean;
-import javax.annotation.Nullable;
+import java.util.function.Consumer;
+import org.eclipse.jdt.annotation.Nullable;
import org.opendaylight.controller.cluster.messaging.MessageAssembler;
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.ApplySnapshot;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow;
import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
* convert to candidate
* </ul>
*/
+// Non-final for testing
public class Follower extends AbstractRaftActorBehavior {
private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
this(context, null, (short)-1);
}
+ @SuppressFBWarnings(value = "MC_OVERRIDABLE_METHOD_CALL_IN_CONSTRUCTOR",
+ justification = "electionDuration() is not final for Candidate override")
public Follower(final RaftActorContext context, final String initialLeaderId,
final short initialLeaderPayloadVersion) {
super(context, RaftState.Follower);
- this.leaderId = initialLeaderId;
- this.leaderPayloadVersion = initialLeaderPayloadVersion;
+ leaderId = initialLeaderId;
+ leaderPayloadVersion = initialLeaderPayloadVersion;
initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), context.getConfigParams()
.getSyncIndexThreshold());
}
@VisibleForTesting
- protected final void setLeaderId(@Nullable final String leaderId) {
+ protected final void setLeaderId(final @Nullable String leaderId) {
this.leaderId = leaderId;
}
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,
- lastIndex(), lastTerm(), context.getPayloadVersion());
+ lastIndex(), lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion());
log.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
sender.tell(reply, actor());
leaderId = appendEntries.getLeaderId();
leaderPayloadVersion = appendEntries.getPayloadVersion();
+ if (appendEntries.getLeaderAddress().isPresent()) {
+ final String address = appendEntries.getLeaderAddress().get();
+ log.debug("New leader address: {}", address);
+
+ context.setPeerAddress(leaderId, address);
+ context.getConfigParams().getPeerAddressResolver().setResolved(leaderId, address);
+ }
+
// First check if the logs are in sync or not
if (isOutOfSync(appendEntries, sender)) {
updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
}
AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
- lastIndex, lastTerm(), context.getPayloadVersion());
+ lastIndex, lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion());
if (log.isTraceEnabled()) {
log.trace("{}: handleAppendEntries returning : {}", logName(), reply);
log.info("{}: Could not remove entries - sending reply to force snapshot", logName());
sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
- lastTerm(), context.getPayloadVersion(), true), actor());
+ lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion()), actor());
return false;
}
break;
} else {
sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
- lastTerm(), context.getPayloadVersion(), true), actor());
+ lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion()), actor());
return false;
}
}
// applied to the state already, as the persistence callback occurs async, and we want those entries
// purged from the persisted log as well.
final AtomicBoolean shouldCaptureSnapshot = new AtomicBoolean(false);
- final Procedure<ReplicatedLogEntry> appendAndPersistCallback = logEntry -> {
+ final Consumer<ReplicatedLogEntry> appendAndPersistCallback = logEntry -> {
final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
final ReplicatedLogEntry lastEntryToAppend = entries.get(entries.size() - 1);
if (shouldCaptureSnapshot.get() && logEntry == lastEntryToAppend) {
log.info("{}: The followers log is empty and the senders prevLogIndex is {}", logName(),
appendEntries.getPrevLogIndex());
- sendOutOfSyncAppendEntriesReply(sender, false);
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
return true;
}
appendEntries.getPrevLogTerm(), lastIndex, context.getReplicatedLog().getSnapshotIndex(),
context.getReplicatedLog().getSnapshotTerm());
- sendOutOfSyncAppendEntriesReply(sender, false);
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
return true;
}
} else if (appendEntries.getPrevLogIndex() != -1) {
+ "snapshotIndex: {}, snapshotTerm: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex,
context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
- sendOutOfSyncAppendEntriesReply(sender, false);
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
return true;
}
}
appendEntries.getReplicatedToAllIndex(), lastIndex,
context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
- sendOutOfSyncAppendEntriesReply(sender, false);
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
return true;
}
entries.get(0).getIndex() - 1, lastIndex, context.getReplicatedLog().getSnapshotIndex(),
context.getReplicatedLog().getSnapshotTerm());
- sendOutOfSyncAppendEntriesReply(sender, false);
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
return true;
}
}
return false;
}
- private void sendOutOfSyncAppendEntriesReply(final ActorRef sender, boolean forceInstallSnapshot) {
+ private void sendOutOfSyncAppendEntriesReply(final ActorRef sender, final boolean forceInstallSnapshot,
+ final short leaderRaftVersion) {
// We found that the log was out of sync so just send a negative reply.
final AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex(),
- lastTerm(), context.getPayloadVersion(), forceInstallSnapshot);
+ lastTerm(), context.getPayloadVersion(), forceInstallSnapshot, needsLeaderAddress(),
+ leaderRaftVersion);
log.info("{}: Follower is out-of-sync so sending negative reply: {}", logName(), reply);
sender.tell(reply, actor());
}
+ private boolean needsLeaderAddress() {
+ return context.getPeerAddress(leaderId) == null;
+ }
+
@Override
protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
final AppendEntriesReply appendEntriesReply) {
return this;
}
+ @Override
+ final ApplyState getApplyStateFor(final ReplicatedLogEntry entry) {
+ return new ApplyState(null, null, entry);
+ }
+
@Override
public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
// If RPC request or response contains term T > currentTerm:
// set currentTerm = T, convert to follower (ยง5.1)
// This applies to all RPC messages and responses
- if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
+ if (rpc.getTerm() > context.getTermInformation().getCurrentTerm() && shouldUpdateTerm(rpc)) {
log.info("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
if (isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
log.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
scheduleElection(electionDuration());
+ } else if (isThisFollowerIsolated()) {
+ log.debug("{}: this follower is isolated. Do not switch to Candidate for now.", logName());
+ setLeaderId(null);
+ scheduleElection(electionDuration());
} else {
log.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
return internalSwitchBehavior(RaftState.Candidate);
return false;
}
+ private boolean isThisFollowerIsolated() {
+ final Optional<Cluster> maybeCluster = context.getCluster();
+ if (!maybeCluster.isPresent()) {
+ return false;
+ }
+
+ final Cluster cluster = maybeCluster.get();
+ final Member selfMember = cluster.selfMember();
+
+ final CurrentClusterState state = cluster.state();
+ final Set<Member> unreachable = state.getUnreachable();
+ final Iterable<Member> members = state.getMembers();
+
+ log.debug("{}: Checking if this node is isolated in the cluster unreachable set {},"
+ + "all members {} self member: {}", logName(), unreachable, members, selfMember);
+
+ // no unreachable peers means we cannot be isolated
+ if (unreachable.size() == 0) {
+ return false;
+ }
+
+ final Set<Member> membersToCheck = new HashSet<>();
+ members.forEach(membersToCheck::add);
+
+ membersToCheck.removeAll(unreachable);
+
+ // check if the only member not unreachable is us
+ if (membersToCheck.size() == 1 && membersToCheck.iterator().next().equals(selfMember)) {
+ return true;
+ }
+
+ return false;
+ }
+
private void handleInstallSnapshot(final ActorRef sender, final InstallSnapshot installSnapshot) {
log.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
installSnapshot.getLastIncludedTerm(),
context.getTermInformation().getCurrentTerm(),
context.getTermInformation().getVotedFor(),
- installSnapshot.getServerConfig().orNull());
+ installSnapshot.getServerConfig().orElse(null));
ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
@Override