import com.google.common.base.Preconditions;
import java.util.ArrayDeque;
import java.util.ArrayList;
+import java.util.Collection;
+import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Queue;
import java.util.UUID;
import javax.annotation.Nullable;
-import org.opendaylight.controller.cluster.raft.RaftActorLeadershipTransferCohort.OnComplete;
import org.opendaylight.controller.cluster.raft.ServerConfigurationPayload.ServerInfo;
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.SnapshotComplete;
import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader;
import org.opendaylight.controller.cluster.raft.messages.AddServer;
import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply;
import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
+import org.opendaylight.yangtools.concepts.Identifier;
+import org.opendaylight.yangtools.util.AbstractUUIDIdentifier;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
+import scala.concurrent.duration.FiniteDuration;
/**
* Handles server configuration related messages for a RaftActor.
}
}
+ void onNewLeader(String leaderId) {
+ currentOperationState.onNewLeader(leaderId);
+ }
+
private void onChangeServersVotingStatus(ChangeServersVotingStatus message, ActorRef sender) {
LOG.debug("{}: onChangeServersVotingStatus: {}, state: {}", raftContext.getId(), message,
currentOperationState);
- onNewOperation(new ChangeServersVotingStatusContext(message, sender));
+ // The following check is a special case. Normally we fail an operation if there's no leader.
+ // Consider a scenario where one has 2 geographically-separated 3-node clusters, one a primary and
+ // the other a backup such that if the primary cluster is lost, the backup can take over. In this
+ // scenario, we have a logical 6-node cluster where the primary sub-cluster is configured as voting
+ // and the backup sub-cluster as non-voting such that the primary cluster can make progress without
+ // consensus from the backup cluster while still replicating to the backup. On fail-over to the backup,
+ // a request would be sent to a member of the backup cluster to flip the voting states, ie make the
+ // backup sub-cluster voting and the lost primary non-voting. However since the primary majority
+ // cluster is lost, there would be no leader to apply, persist and replicate the server config change.
+ // Therefore, if the local server is currently non-voting and is to be changed to voting and there is
+ // no current leader, we will try to elect a leader using the new server config in order to replicate
+ // the change and progress.
+ boolean localServerChangingToVoting = Boolean.TRUE.equals(message.
+ getServerVotingStatusMap().get(raftActor.getRaftActorContext().getId()));
+ boolean hasNoLeader = raftActor.getLeaderId() == null;
+ if(localServerChangingToVoting && !raftContext.isVotingMember() && hasNoLeader) {
+ currentOperationState.onNewOperation(new ChangeServersVotingStatusContext(message, sender, true));
+ } else {
+ onNewOperation(new ChangeServersVotingStatusContext(message, sender, false));
+ }
}
private void onRemoveServer(RemoveServer removeServer, ActorRef sender) {
LOG.debug("{}: onRemoveServer: {}, state: {}", raftContext.getId(), removeServer, currentOperationState);
- boolean isSelf = removeServer.getServerId().equals(raftActor.getId());
+ boolean isSelf = removeServer.getServerId().equals(raftContext.getId());
if(isSelf && !raftContext.hasFollowers()) {
sender.tell(new RemoveServerReply(ServerChangeStatus.NOT_SUPPORTED, raftActor.getLeaderId()),
raftActor.getSelf());
ActorSelection leader = raftActor.getLeader();
if (leader != null) {
LOG.debug("{}: Not leader - forwarding to leader {}", raftContext.getId(), leader);
- leader.forward(operationContext.getOperation(), raftActor.getContext());
+ leader.tell(operationContext.getOperation(), operationContext.getClientRequestor());
} else {
LOG.debug("{}: No leader - returning NO_LEADER reply", raftContext.getId());
operationContext.getClientRequestor().tell(operationContext.newReply(
}
+ void onNewLeader(String newLeader) {
+ }
+
protected void persistNewServerConfiguration(ServerOperationContext<?> operationContext){
raftContext.setDynamicServerConfigurationInUse();
operationContext.operationComplete(raftActor, replyStatus == null || replyStatus == ServerChangeStatus.OK);
+ changeToIdleState();
+ }
+
+ protected void changeToIdleState() {
currentOperationState = IDLE;
ServerOperationContext<?> nextOperation = pendingOperationsQueue.poll();
}
Cancellable newTimer(Object message) {
+ return newTimer(raftContext.getConfigParams().getElectionTimeOutInterval().$times(2), message);
+ }
+
+ Cancellable newTimer(FiniteDuration timeout, Object message) {
return raftContext.getActorSystem().scheduler().scheduleOnce(
- raftContext.getConfigParams().getElectionTimeOutInterval().$times(2), raftContext.getActor(), message,
+ timeout, raftContext.getActor(), message,
raftContext.getActorSystem().dispatcher(), raftContext.getActor());
}
// Sanity check - we could get an ApplyState from a previous operation that timed out so make
// sure it's meant for us.
if(operationContext.getContextId().equals(applyState.getIdentifier())) {
- LOG.info("{}: {} has been successfully replicated to a majority of followers", raftActor.getId(),
+ LOG.info("{}: {} has been successfully replicated to a majority of followers", raftContext.getId(),
applyState.getReplicatedLogEntry().getData());
timer.cancel();
}
}
+ private static final class ServerOperationContextIdentifier extends AbstractUUIDIdentifier<ServerOperationContextIdentifier> {
+ private static final long serialVersionUID = 1L;
+
+ ServerOperationContextIdentifier() {
+ super(UUID.randomUUID());
+ }
+ }
+
/**
* Stores context information for a server operation.
*
private static abstract class ServerOperationContext<T> {
private final T operation;
private final ActorRef clientRequestor;
- private final String contextId;
+ private final Identifier contextId;
ServerOperationContext(T operation, ActorRef clientRequestor){
this.operation = operation;
this.clientRequestor = clientRequestor;
- contextId = UUID.randomUUID().toString();
+ contextId = new ServerOperationContextIdentifier();
}
- String getContextId() {
+ Identifier getContextId() {
return contextId;
}
}
private static class ChangeServersVotingStatusContext extends ServerOperationContext<ChangeServersVotingStatus> {
- ChangeServersVotingStatusContext(ChangeServersVotingStatus convertMessage, ActorRef clientRequestor) {
+ private final boolean tryToElectLeader;
+
+ ChangeServersVotingStatusContext(ChangeServersVotingStatus convertMessage, ActorRef clientRequestor,
+ boolean tryToElectLeader) {
super(convertMessage, clientRequestor);
+ this.tryToElectLeader = tryToElectLeader;
}
@Override
InitialOperationState newInitialOperationState(RaftActorServerConfigurationSupport support) {
- return support.new ChangeServersVotingStatusState(this);
+ return support.new ChangeServersVotingStatusState(this, tryToElectLeader);
}
@Override
// leadership.
boolean localServerChangedToNonVoting = Boolean.FALSE.equals(getOperation().
getServerVotingStatusMap().get(raftActor.getRaftActorContext().getId()));
- if(succeeded && localServerChangedToNonVoting && raftActor.isLeader()) {
- raftActor.initiateLeadershipTransfer(new OnComplete() {
- @Override
- public void onSuccess(ActorRef raftActorRef, ActorRef replyTo) {
- LOG.debug("{}: leader transfer succeeded after change to non-voting", raftActor.persistenceId());
- ensureFollowerState(raftActor);
- }
-
- @Override
- public void onFailure(ActorRef raftActorRef, ActorRef replyTo) {
- LOG.debug("{}: leader transfer failed after change to non-voting", raftActor.persistenceId());
- ensureFollowerState(raftActor);
- }
-
- private void ensureFollowerState(RaftActor raftActor) {
- // Whether or not leadership transfer succeeded, we have to step down as leader and
- // switch to Follower so ensure that.
- if(raftActor.getRaftState() != RaftState.Follower) {
- raftActor.initializeBehavior();
- }
- }
- });
+ if (succeeded && localServerChangedToNonVoting) {
+ LOG.debug("Leader changed to non-voting - trying leadership transfer");
+ raftActor.becomeNonVoting();
}
}
@Override
String getLoggingContext() {
- return getOperation().getServerVotingStatusMap().toString();
+ return getOperation().toString();
}
}
private class ChangeServersVotingStatusState extends OperationState implements InitialOperationState {
private final ChangeServersVotingStatusContext changeVotingStatusContext;
+ private final boolean tryToElectLeader;
- ChangeServersVotingStatusState(ChangeServersVotingStatusContext changeVotingStatusContext) {
+ ChangeServersVotingStatusState(ChangeServersVotingStatusContext changeVotingStatusContext,
+ boolean tryToElectLeader) {
this.changeVotingStatusContext = changeVotingStatusContext;
+ this.tryToElectLeader = tryToElectLeader;
}
@Override
public void initiate() {
LOG.debug("Initiating ChangeServersVotingStatusState");
+ if(tryToElectLeader) {
+ initiateLocalLeaderElection();
+ } else if(updateLocalPeerInfo()) {
+ persistNewServerConfiguration(changeVotingStatusContext);
+ }
+ }
+
+ private void initiateLocalLeaderElection() {
+ LOG.debug("{}: Sending local ElectionTimeout to start leader election", raftContext.getId());
+
+ ServerConfigurationPayload previousServerConfig = raftContext.getPeerServerInfo(true);
+ if(!updateLocalPeerInfo()) {
+ return;
+ }
+
+ raftContext.getActor().tell(ElectionTimeout.INSTANCE, raftContext.getActor());
+
+ currentOperationState = new WaitingForLeaderElected(changeVotingStatusContext, previousServerConfig);
+ }
+
+ private boolean updateLocalPeerInfo() {
+ List<ServerInfo> newServerInfoList = newServerInfoList();
+
+ // Check if new voting state would leave us with no voting members.
+ boolean atLeastOneVoting = false;
+ for(ServerInfo info: newServerInfoList) {
+ if(info.isVoting()) {
+ atLeastOneVoting = true;
+ break;
+ }
+ }
+
+ if(!atLeastOneVoting) {
+ operationComplete(changeVotingStatusContext, ServerChangeStatus.INVALID_REQUEST);
+ return false;
+ }
+
+ raftContext.updatePeerIds(new ServerConfigurationPayload(newServerInfoList));
+ if(raftActor.getCurrentBehavior() instanceof AbstractLeader) {
+ AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior();
+ leader.updateMinReplicaCount();
+ }
+
+ return true;
+ }
+
+ private List<ServerInfo> newServerInfoList() {
Map<String, Boolean> serverVotingStatusMap = changeVotingStatusContext.getOperation().getServerVotingStatusMap();
List<ServerInfo> newServerInfoList = new ArrayList<>();
for(String peerId: raftContext.getPeerIds()) {
newServerInfoList.add(new ServerInfo(raftContext.getId(), serverVotingStatusMap.containsKey(
raftContext.getId()) ? serverVotingStatusMap.get(raftContext.getId()) : raftContext.isVotingMember()));
- raftContext.updatePeerIds(new ServerConfigurationPayload(newServerInfoList));
- AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior();
- leader.updateMinReplicaCount();
+ return newServerInfoList;
+ }
+ }
+
+ private class WaitingForLeaderElected extends OperationState {
+ private final ServerConfigurationPayload previousServerConfig;
+ private final ChangeServersVotingStatusContext operationContext;
+ private final Cancellable timer;
+
+ WaitingForLeaderElected(ChangeServersVotingStatusContext operationContext,
+ ServerConfigurationPayload previousServerConfig) {
+ this.operationContext = operationContext;
+ this.previousServerConfig = previousServerConfig;
+
+ timer = newTimer(raftContext.getConfigParams().getElectionTimeOutInterval(),
+ new ServerOperationTimeout(operationContext.getLoggingContext()));
+ }
+
+ @Override
+ void onNewLeader(String newLeader) {
+ LOG.debug("{}: New leader {} elected", raftContext.getId(), newLeader);
+
+ timer.cancel();
+
+ if(raftActor.isLeader()) {
+ persistNewServerConfiguration(operationContext);
+ } else {
+ // Edge case - some other node became leader so forward the operation.
+ LOG.debug("{}: Forwarding {} to new leader", raftContext.getId(), operationContext.getOperation());
+
+ // Revert the local server config change.
+ raftContext.updatePeerIds(previousServerConfig);
- persistNewServerConfiguration(changeVotingStatusContext);
+ changeToIdleState();
+ RaftActorServerConfigurationSupport.this.onNewOperation(operationContext);
+ }
+ }
+
+ @Override
+ void onServerOperationTimeout(ServerOperationTimeout timeout) {
+ LOG.warn("{}: Leader election timed out - cannot apply operation {}",
+ raftContext.getId(), timeout.getLoggingContext());
+
+ // Revert the local server config change.
+ raftContext.updatePeerIds(previousServerConfig);
+ raftActor.initializeBehavior();
+
+ tryToForwardOperationToAnotherServer();
+ }
+
+ private void tryToForwardOperationToAnotherServer() {
+ Collection<String> serversVisited = new HashSet<>(operationContext.getOperation().getServersVisited());
+
+ LOG.debug("{}: tryToForwardOperationToAnotherServer - servers already visited {}", raftContext.getId(),
+ serversVisited);
+
+ serversVisited.add(raftContext.getId());
+
+ // Try to find another whose state is being changed from non-voting to voting and that we haven't
+ // tried yet.
+ Map<String, Boolean> serverVotingStatusMap = operationContext.getOperation().getServerVotingStatusMap();
+ ActorSelection forwardToPeerActor = null;
+ for(Map.Entry<String, Boolean> e: serverVotingStatusMap.entrySet()) {
+ Boolean isVoting = e.getValue();
+ String serverId = e.getKey();
+ PeerInfo peerInfo = raftContext.getPeerInfo(serverId);
+ if(isVoting && peerInfo != null && !peerInfo.isVoting() && !serversVisited.contains(serverId)) {
+ ActorSelection actor = raftContext.getPeerActorSelection(serverId);
+ if(actor != null) {
+ forwardToPeerActor = actor;
+ break;
+ }
+ }
+ }
+
+ if(forwardToPeerActor != null) {
+ LOG.debug("{}: Found server {} to forward to", raftContext.getId(), forwardToPeerActor);
+
+ forwardToPeerActor.tell(new ChangeServersVotingStatus(serverVotingStatusMap, serversVisited),
+ operationContext.getClientRequestor());
+ changeToIdleState();
+ } else {
+ operationComplete(operationContext, ServerChangeStatus.NO_LEADER);
+ }
}
}
Code Review / controller.git / blobdiff