+ /**
+ * Interface for the initial state for a server operation.
+ */
+ private interface InitialOperationState {
+ void initiate();
+ }
+
+ /**
+ * Abstract base class for a server operation FSM state. Handles common behavior for all states.
+ */
+ private abstract class OperationState {
+ void onNewOperation(ServerOperationContext<?> operationContext) {
+ // We're currently processing another operation so queue it to be processed later.
+
+ LOG.debug("{}: Server operation already in progress - queueing {}", raftContext.getId(),
+ operationContext.getOperation());
+
+ pendingOperationsQueue.add(operationContext);
+ }
+
+ void onServerOperationTimeout(ServerOperationTimeout timeout) {
+ LOG.debug("onServerOperationTimeout should not be called in state {}", this);
+ }
+
+ void onUnInitializedFollowerSnapshotReply(UnInitializedFollowerSnapshotReply reply) {
+ LOG.debug("onUnInitializedFollowerSnapshotReply was called in state {}", this);
+ }
+
+ void onApplyState(ApplyState applyState) {
+ LOG.debug("onApplyState was called in state {}", this);
+ }
+
+ void onSnapshotComplete() {
+
+ }
+
+ void onNewLeader(String newLeader) {
+ }
+
+ protected void persistNewServerConfiguration(ServerOperationContext<?> operationContext) {
+ raftContext.setDynamicServerConfigurationInUse();
+
+ ServerConfigurationPayload payload = raftContext.getPeerServerInfo(
+ operationContext.includeSelfInNewConfiguration(raftActor));
+ LOG.debug("{}: New server configuration : {}", raftContext.getId(), payload.getServerConfig());
+
+ raftActor.persistData(operationContext.getClientRequestor(), operationContext.getContextId(),
+ payload, false);
+
+ currentOperationState = new Persisting(operationContext, newTimer(new ServerOperationTimeout(
+ operationContext.getLoggingContext())));
+
+ sendReply(operationContext, ServerChangeStatus.OK);
+ }
+
+ protected void operationComplete(ServerOperationContext<?> operationContext,
+ @Nullable ServerChangeStatus replyStatus) {
+ if (replyStatus != null) {
+ sendReply(operationContext, replyStatus);
+ }
+
+ operationContext.operationComplete(raftActor, replyStatus == null || replyStatus == ServerChangeStatus.OK);
+
+ changeToIdleState();
+ }
+
+ protected void changeToIdleState() {
+ currentOperationState = IDLE;
+
+ ServerOperationContext<?> nextOperation = pendingOperationsQueue.poll();
+ if (nextOperation != null) {
+ RaftActorServerConfigurationSupport.this.onNewOperation(nextOperation);
+ }
+ }
+
+ protected void sendReply(ServerOperationContext<?> operationContext, ServerChangeStatus status) {
+ LOG.debug("{}: Returning {} for operation {}", raftContext.getId(), status,
+ operationContext.getOperation());
+
+ operationContext.getClientRequestor().tell(operationContext.newReply(status, raftActor.getLeaderId()),
+ raftActor.self());
+ }
+
+ Cancellable newTimer(Object message) {
+ return newTimer(raftContext.getConfigParams().getElectionTimeOutInterval().$times(2), message);
+ }
+
+ Cancellable newTimer(FiniteDuration timeout, Object message) {
+ return raftContext.getActorSystem().scheduler().scheduleOnce(
+ timeout, raftContext.getActor(), message,
+ raftContext.getActorSystem().dispatcher(), raftContext.getActor());
+ }
+
+ @Override
+ public String toString() {
+ return getClass().getSimpleName();
+ }
+ }
+
+ /**
+ * The state when no server operation is in progress. It immediately initiates new server operations.
+ */
+ private final class Idle extends OperationState {
+ @Override
+ public void onNewOperation(ServerOperationContext<?> operationContext) {
+ operationContext.newInitialOperationState(RaftActorServerConfigurationSupport.this).initiate();
+ }
+
+ @Override
+ public void onApplyState(ApplyState applyState) {
+ // Noop - we override b/c ApplyState is called normally for followers in the idle state.
+ }
+ }
+
+ /**
+ * The state when a new server configuration is being persisted and replicated.
+ */
+ private final class Persisting extends OperationState {
+ private final ServerOperationContext<?> operationContext;
+ private final Cancellable timer;
+ private boolean timedOut = false;
+
+ Persisting(ServerOperationContext<?> operationContext, Cancellable timer) {
+ this.operationContext = operationContext;
+ this.timer = timer;
+ }
+
+ @Override
+ public void onApplyState(ApplyState applyState) {
+ // 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", raftContext.getId(),
+ applyState.getReplicatedLogEntry().getData());
+
+ timer.cancel();
+ operationComplete(operationContext, null);
+ }
+ }
+
+ @Override
+ public void onServerOperationTimeout(ServerOperationTimeout timeout) {
+ LOG.warn("{}: Timeout occured while replicating the new server configuration for {}", raftContext.getId(),
+ timeout.getLoggingContext());
+
+ timedOut = true;
+
+ // Fail any pending operations
+ ServerOperationContext<?> nextOperation = pendingOperationsQueue.poll();
+ while (nextOperation != null) {
+ sendReply(nextOperation, ServerChangeStatus.PRIOR_REQUEST_CONSENSUS_TIMEOUT);
+ nextOperation = pendingOperationsQueue.poll();
+ }
+ }
+
+ @Override
+ public void onNewOperation(ServerOperationContext<?> newOperationContext) {
+ if (timedOut) {
+ sendReply(newOperationContext, ServerChangeStatus.PRIOR_REQUEST_CONSENSUS_TIMEOUT);
+ } else {
+ super.onNewOperation(newOperationContext);
+ }
+ }
+ }
+
+ /**
+ * Abstract base class for an AddServer operation state.
+ */
+ private abstract class AddServerState extends OperationState {
+ private final AddServerContext addServerContext;
+
+ AddServerState(AddServerContext addServerContext) {
+ this.addServerContext = addServerContext;
+ }
+
+ AddServerContext getAddServerContext() {
+ return addServerContext;
+ }
+
+ Cancellable newInstallSnapshotTimer() {
+ return newTimer(new ServerOperationTimeout(addServerContext.getOperation().getNewServerId()));
+ }
+
+ void handleInstallSnapshotTimeout(ServerOperationTimeout timeout) {
+ String serverId = timeout.getLoggingContext();
+
+ LOG.debug("{}: handleInstallSnapshotTimeout for new server {}", raftContext.getId(), serverId);
+
+ // cleanup
+ raftContext.removePeer(serverId);
+
+ boolean isLeader = raftActor.isLeader();
+ if (isLeader) {
+ AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior();
+ leader.removeFollower(serverId);
+ }
+
+ operationComplete(getAddServerContext(), isLeader ? ServerChangeStatus.TIMEOUT
+ : ServerChangeStatus.NO_LEADER);
+ }
+
+ }
+
+ /**
+ * The initial state for the AddServer operation. It adds the new follower as a peer and initiates
+ * snapshot capture, if necessary.
+ */
+ private final class InitialAddServerState extends AddServerState implements InitialOperationState {
+ InitialAddServerState(AddServerContext addServerContext) {
+ super(addServerContext);
+ }
+
+ @Override
+ public void initiate() {
+ final AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior();
+ AddServer addServer = getAddServerContext().getOperation();
+
+ LOG.debug("{}: Initiating {}", raftContext.getId(), addServer);
+
+ if (raftContext.getPeerInfo(addServer.getNewServerId()) != null) {
+ operationComplete(getAddServerContext(), ServerChangeStatus.ALREADY_EXISTS);
+ return;
+ }
+
+ VotingState votingState = addServer.isVotingMember() ? VotingState.VOTING_NOT_INITIALIZED :
+ VotingState.NON_VOTING;
+ raftContext.addToPeers(addServer.getNewServerId(), addServer.getNewServerAddress(), votingState);
+
+ leader.addFollower(addServer.getNewServerId());
+
+ if (votingState == VotingState.VOTING_NOT_INITIALIZED) {
+ // schedule the install snapshot timeout timer
+ Cancellable installSnapshotTimer = newInstallSnapshotTimer();
+ if (leader.initiateCaptureSnapshot(addServer.getNewServerId())) {
+ LOG.debug("{}: Initiating capture snapshot for new server {}", raftContext.getId(),
+ addServer.getNewServerId());
+
+ currentOperationState = new InstallingSnapshot(getAddServerContext(), installSnapshotTimer);
+ } else {
+ LOG.debug("{}: Snapshot already in progress - waiting for completion", raftContext.getId());
+
+ currentOperationState = new WaitingForPriorSnapshotComplete(getAddServerContext(),
+ installSnapshotTimer);
+ }
+ } else {
+ LOG.debug("{}: New follower is non-voting - directly persisting new server configuration",
+ raftContext.getId());
+
+ persistNewServerConfiguration(getAddServerContext());
+ }
+ }
+ }
+
+ /**
+ * The AddServer operation state for when the catch-up snapshot is being installed. It handles successful
+ * reply or timeout.
+ */
+ private final class InstallingSnapshot extends AddServerState {
+ private final Cancellable installSnapshotTimer;
+
+ InstallingSnapshot(AddServerContext addServerContext, Cancellable installSnapshotTimer) {
+ super(addServerContext);
+ this.installSnapshotTimer = Preconditions.checkNotNull(installSnapshotTimer);
+ }
+
+ @Override
+ public void onServerOperationTimeout(ServerOperationTimeout timeout) {
+ handleInstallSnapshotTimeout(timeout);
+
+ LOG.warn("{}: Timeout occured for new server {} while installing snapshot", raftContext.getId(),
+ timeout.getLoggingContext());
+ }
+
+ @Override
+ public void onUnInitializedFollowerSnapshotReply(UnInitializedFollowerSnapshotReply reply) {
+ LOG.debug("{}: onUnInitializedFollowerSnapshotReply: {}", raftContext.getId(), reply);
+
+ String followerId = reply.getFollowerId();
+
+ // Sanity check to guard against receiving an UnInitializedFollowerSnapshotReply from a prior
+ // add server operation that timed out.
+ if (getAddServerContext().getOperation().getNewServerId().equals(followerId) && raftActor.isLeader()) {
+ AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior();
+ raftContext.getPeerInfo(followerId).setVotingState(VotingState.VOTING);
+ leader.updateMinReplicaCount();
+
+ persistNewServerConfiguration(getAddServerContext());
+
+ installSnapshotTimer.cancel();
+ } else {
+ LOG.debug("{}: Dropping UnInitializedFollowerSnapshotReply for server {}: {}",
+ raftContext.getId(), followerId,
+ !raftActor.isLeader() ? "not leader" : "server Id doesn't match");
+ }
+ }
+ }
+
+ /**
+ * The AddServer operation state for when there is a snapshot already in progress. When the current
+ * snapshot completes, it initiates an install snapshot.
+ */
+ private final class WaitingForPriorSnapshotComplete extends AddServerState {
+ private final Cancellable snapshotTimer;
+
+ WaitingForPriorSnapshotComplete(AddServerContext addServerContext, Cancellable snapshotTimer) {
+ super(addServerContext);
+ this.snapshotTimer = Preconditions.checkNotNull(snapshotTimer);
+ }
+
+ @Override
+ public void onSnapshotComplete() {
+ LOG.debug("{}: onSnapshotComplete", raftContext.getId());
+
+ if (!raftActor.isLeader()) {
+ LOG.debug("{}: No longer the leader", raftContext.getId());
+ return;
+ }
+
+ AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior();
+ if (leader.initiateCaptureSnapshot(getAddServerContext().getOperation().getNewServerId())) {
+ LOG.debug("{}: Initiating capture snapshot for new server {}", raftContext.getId(),
+ getAddServerContext().getOperation().getNewServerId());
+
+ currentOperationState = new InstallingSnapshot(getAddServerContext(),
+ newInstallSnapshotTimer());
+
+ snapshotTimer.cancel();
+ }
+ }
+
+ @Override
+ public void onServerOperationTimeout(ServerOperationTimeout timeout) {
+ handleInstallSnapshotTimeout(timeout);
+
+ LOG.warn("{}: Timeout occured for new server {} while waiting for prior snapshot to complete",
+ raftContext.getId(), timeout.getLoggingContext());
+ }
+ }
+
+ 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.
+ *
+ * @param <T> the operation type
+ */
+ private abstract static class ServerOperationContext<T> {
+ private final T operation;
+ private final ActorRef clientRequestor;
+ private final Identifier contextId;
+
+ ServerOperationContext(T operation, ActorRef clientRequestor) {
+ this.operation = operation;
+ this.clientRequestor = clientRequestor;
+ contextId = new ServerOperationContextIdentifier();
+ }
+
+ Identifier getContextId() {
+ return contextId;
+ }
+
+ T getOperation() {
+ return operation;
+ }
+
+ ActorRef getClientRequestor() {
+ return clientRequestor;
+ }
+
+ void operationComplete(RaftActor raftActor, boolean succeeded) {
+ }
+
+ boolean includeSelfInNewConfiguration(RaftActor raftActor) {
+ return true;
+ }
+
+ abstract Object newReply(ServerChangeStatus status, String leaderId);
+
+ abstract InitialOperationState newInitialOperationState(RaftActorServerConfigurationSupport support);
+
+ abstract String getLoggingContext();
+ }
+
+ /**
+ * Stores context information for an AddServer operation.
+ */
+ private static class AddServerContext extends ServerOperationContext<AddServer> {
+ AddServerContext(AddServer addServer, ActorRef clientRequestor) {
+ super(addServer, clientRequestor);
+ }
+
+ @Override
+ Object newReply(ServerChangeStatus status, String leaderId) {
+ return new AddServerReply(status, leaderId);