import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
-import java.util.Iterator;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;
import org.opendaylight.controller.cluster.messaging.MessageSlicer;
import org.opendaylight.controller.cluster.messaging.SliceOptions;
import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
-import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
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.RaftVersions;
import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
import org.opendaylight.controller.cluster.raft.VotingState;
+import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
import org.opendaylight.controller.cluster.raft.base.messages.CheckConsensusReached;
import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
+import org.opendaylight.controller.cluster.raft.messages.IdentifiablePayload;
import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
followerToLog.remove(followerId);
}
- public void updateMinReplicaCount() {
+ public final void updateMinReplicaCount() {
int numVoting = 0;
for (PeerInfo peer: context.getPeers()) {
if (peer.isVoting()) {
return this;
}
+ final var followerRaftVersion = appendEntriesReply.getRaftVersion();
+ if (followerRaftVersion < RaftVersions.FLUORINE_VERSION) {
+ log.warn("{}: handleAppendEntriesReply - ignoring reply from follower {} raft version {}", logName(),
+ followerId, followerRaftVersion);
+ return this;
+ }
+
final long lastActivityNanos = followerLogInformation.nanosSinceLastActivity();
if (lastActivityNanos > context.getConfigParams().getElectionTimeOutInterval().toNanos()) {
log.warn("{} : handleAppendEntriesReply delayed beyond election timeout, "
followerLogInformation.markFollowerActive();
followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
- followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
+ followerLogInformation.setRaftVersion(followerRaftVersion);
followerLogInformation.setNeedsLeaderAddress(appendEntriesReply.isNeedsLeaderAddress());
long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
}
- @Override
- protected ClientRequestTracker removeClientRequestTracker(final long logIndex) {
- final Iterator<ClientRequestTracker> it = trackers.iterator();
+ /**
+ * Removes and returns the ClientRequestTracker for the specified log index.
+ * @param logIndex the log index
+ * @return the ClientRequestTracker or null if none available
+ */
+ private ClientRequestTracker removeClientRequestTracker(final long logIndex) {
+ final var it = trackers.iterator();
while (it.hasNext()) {
- final ClientRequestTracker t = it.next();
- if (t.getIndex() == logIndex) {
+ final var tracker = it.next();
+ if (tracker.logIndex() == logIndex) {
it.remove();
- return t;
+ return tracker;
}
}
-
return null;
}
@Override
- protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
- final RequestVoteReply requestVoteReply) {
+ final ApplyState getApplyStateFor(final ReplicatedLogEntry entry) {
+ // first check whether a ClientRequestTracker exists for this entry.
+ // If it does that means the leader wasn't dropped before the transaction applied.
+ // That means that this transaction can be safely applied as a local transaction since we
+ // have the ClientRequestTracker.
+ final var tracker = removeClientRequestTracker(entry.getIndex());
+ if (tracker != null) {
+ return new ApplyState(tracker.clientActor(), tracker.identifier(), entry);
+ }
+
+ // Tracker is missing, this means that we switched behaviours between replicate and applystate
+ // and became the leader again,. We still want to apply this as a local modification because
+ // we have resumed leadership with that log entry having been committed.
+ if (entry.getData() instanceof IdentifiablePayload<?> identifiable) {
+ return new ApplyState(null, identifiable.getIdentifier(), entry);
+ }
+
+ return new ApplyState(null, null, entry);
+ }
+
+ @Override
+ protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender, final RequestVoteReply requestVoteReply) {
return this;
}
- protected void beforeSendHeartbeat(){}
+ protected void beforeSendHeartbeat() {
+ // No-op
+ }
@Override
public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
return this;
}
- if (message instanceof RaftRPC) {
- RaftRPC rpc = (RaftRPC) message;
- // 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 leader's term {} - switching to Follower",
- logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
-
- context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
-
- // This is a special case. Normally when stepping down as leader we don't process and reply to the
- // RaftRPC as per raft. But if we're in the process of transferring leadership and we get a
- // RequestVote, process the RequestVote before switching to Follower. This enables the requesting
- // candidate node to be elected the leader faster and avoids us possibly timing out in the Follower
- // state and starting a new election and grabbing leadership back before the other candidate node can
- // start a new election due to lack of responses. This case would only occur if there isn't a majority
- // of other nodes available that can elect the requesting candidate. Since we're transferring
- // leadership, we should make every effort to get the requesting node elected.
- if (message instanceof RequestVote && context.getRaftActorLeadershipTransferCohort() != null) {
- log.debug("{}: Leadership transfer in progress - processing RequestVote", logName());
- super.handleMessage(sender, message);
- }
-
- return internalSwitchBehavior(RaftState.Follower);
+ // 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 (message instanceof RaftRPC rpc && rpc.getTerm() > context.getTermInformation().getCurrentTerm()
+ && shouldUpdateTerm(rpc)) {
+
+ log.info("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
+ logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
+
+ context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
+
+ // This is a special case. Normally when stepping down as leader we don't process and reply to the
+ // RaftRPC as per raft. But if we're in the process of transferring leadership and we get a
+ // RequestVote, process the RequestVote before switching to Follower. This enables the requesting
+ // candidate node to be elected the leader faster and avoids us possibly timing out in the Follower
+ // state and starting a new election and grabbing leadership back before the other candidate node can
+ // start a new election due to lack of responses. This case would only occur if there isn't a majority
+ // of other nodes available that can elect the requesting candidate. Since we're transferring
+ // leadership, we should make every effort to get the requesting node elected.
+ if (rpc instanceof RequestVote requestVote && context.getRaftActorLeadershipTransferCohort() != null) {
+ log.debug("{}: Leadership transfer in progress - processing RequestVote", logName());
+ requestVote(sender, requestVote);
}
+
+ return internalSwitchBehavior(RaftState.Follower);
}
if (message instanceof SendHeartBeat) {
beforeSendHeartbeat();
sendHeartBeat();
scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
- } else if (message instanceof SendInstallSnapshot) {
- SendInstallSnapshot sendInstallSnapshot = (SendInstallSnapshot) message;
+ } else if (message instanceof SendInstallSnapshot sendInstallSnapshot) {
setSnapshotHolder(new SnapshotHolder(sendInstallSnapshot.getSnapshot(),
sendInstallSnapshot.getSnapshotBytes()));
sendInstallSnapshot();
- } else if (message instanceof Replicate) {
- replicate((Replicate) message);
- } else if (message instanceof InstallSnapshotReply) {
- handleInstallSnapshotReply((InstallSnapshotReply) message);
+ } else if (message instanceof Replicate replicate) {
+ replicate(replicate);
+ } else if (message instanceof InstallSnapshotReply installSnapshotReply) {
+ handleInstallSnapshotReply(installSnapshotReply);
} else if (message instanceof CheckConsensusReached) {
possiblyUpdateCommitIndex();
} else {
if (installSnapshotState.isLastChunk(reply.getChunkIndex())) {
//this was the last chunk reply
- long followerMatchIndex = snapshotHolder.get().getLastIncludedIndex();
+ long followerMatchIndex = snapshotHolder.orElseThrow().getLastIncludedIndex();
followerLogInformation.setMatchIndex(followerMatchIndex);
followerLogInformation.setNextIndex(followerMatchIndex + 1);
followerLogInformation.clearLeaderInstallSnapshotState();
}
private void replicate(final Replicate replicate) {
- long logIndex = replicate.getReplicatedLogEntry().getIndex();
+ final long logIndex = replicate.logIndex();
- log.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}, isSendImmediate: {}", logName(),
- replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass(),
- replicate.isSendImmediate());
+ log.debug("{}: Replicate message: identifier: {}, logIndex: {}, isSendImmediate: {}", logName(),
+ replicate.identifier(), logIndex, replicate.sendImmediate());
// Create a tracker entry we will use this later to notify the
// client actor
- if (replicate.getClientActor() != null) {
- trackers.add(new ClientRequestTrackerImpl(replicate.getClientActor(), replicate.getIdentifier(),
- logIndex));
+ final var clientActor = replicate.clientActor();
+ if (clientActor != null) {
+ trackers.add(new ClientRequestTracker(logIndex, clientActor, replicate.identifier()));
}
boolean applyModificationToState = !context.anyVotingPeers()
applyLogToStateMachine(logIndex);
}
- if (replicate.isSendImmediate() && !followerToLog.isEmpty()) {
+ if (replicate.sendImmediate() && !followerToLog.isEmpty()) {
sendAppendEntries(0, false);
}
}
// If the first entry's size exceeds the max data size threshold, it will be returned from the call above. If
// that is the case, then we need to slice it into smaller chunks.
- if (!(entries.size() == 1 && entries.get(0).getData().size() > maxDataSize)) {
+ if (entries.size() != 1 || entries.get(0).getData().serializedSize() <= maxDataSize) {
// Don't need to slice.
return entries;
}
}
boolean captureInitiated = context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
- this.getReplicatedToAllIndex(), followerId);
+ getReplicatedToAllIndex(), followerId);
if (captureInitiated) {
followerLogInfo.setLeaderInstallSnapshotState(new LeaderInstallSnapshotState(
context.getConfigParams().getSnapshotChunkSize(), logName()));
try {
// Ensure the snapshot bytes are set - this is a no-op.
- installSnapshotState.setSnapshotBytes(snapshotHolder.get().getSnapshotBytes());
+ installSnapshotState.setSnapshotBytes(snapshotHolder.orElseThrow().getSnapshotBytes());
if (!installSnapshotState.canSendNextChunk()) {
return;
} catch (IOException e) {
log.warn("{}: Unable to send chunk: {}/{}. Reseting snapshot progress. Snapshot state: {}", logName(),
installSnapshotState.getChunkIndex(), installSnapshotState.getTotalChunks(),
- installSnapshotState);
+ installSnapshotState, e);
installSnapshotState.reset();
}
}
installSnapshotState.startChunkTimer();
followerActor.tell(
new InstallSnapshot(currentTerm(), context.getId(),
- snapshotHolder.get().getLastIncludedIndex(),
- snapshotHolder.get().getLastIncludedTerm(),
+ snapshotHolder.orElseThrow().getLastIncludedIndex(),
+ snapshotHolder.orElseThrow().getLastIncludedTerm(),
snapshotChunk,
chunkIndex,
installSnapshotState.getTotalChunks(),
OptionalInt.of(installSnapshotState.getLastChunkHashCode()),
- serverConfig
- ).toSerializable(followerLogInfo.getRaftVersion()),
+ serverConfig,
+ followerLogInfo.getRaftVersion()),
actor()
);
}
private final ByteSource snapshotBytes;
SnapshotHolder(final Snapshot snapshot, final ByteSource snapshotBytes) {
- this.lastIncludedTerm = snapshot.getLastAppliedTerm();
- this.lastIncludedIndex = snapshot.getLastAppliedIndex();
+ lastIncludedTerm = snapshot.getLastAppliedTerm();
+ lastIncludedIndex = snapshot.getLastAppliedIndex();
this.snapshotBytes = snapshotBytes;
}