X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fsal-akka-raft%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fraft%2FRaftActorServerConfigurationSupport.java;h=29641cb00e441d5dedc5e7cccccd69bb0b0e6d3a;hb=refs%2Fchanges%2F70%2F82770%2F4;hp=ae23140114ed7d2ca8e73584ecd997fefd437723;hpb=3fda1a923defdbf18849c6080c3aa19f1ebf2c5f;p=controller.git diff --git a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActorServerConfigurationSupport.java b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActorServerConfigurationSupport.java index ae23140114..29641cb00e 100644 --- a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActorServerConfigurationSupport.java +++ b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/RaftActorServerConfigurationSupport.java @@ -7,25 +7,38 @@ */ package org.opendaylight.controller.cluster.raft; +import static java.util.Objects.requireNonNull; + import akka.actor.ActorRef; import akka.actor.ActorSelection; import akka.actor.Cancellable; +import java.util.ArrayDeque; import java.util.ArrayList; -import java.util.Collections; -import java.util.LinkedList; +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 java.util.concurrent.TimeUnit; -import org.opendaylight.controller.cluster.raft.FollowerLogInformation.FollowerState; +import org.eclipse.jdt.annotation.Nullable; import org.opendaylight.controller.cluster.raft.base.messages.ApplyState; +import org.opendaylight.controller.cluster.raft.base.messages.SnapshotComplete; +import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow; import org.opendaylight.controller.cluster.raft.behaviors.AbstractLeader; import org.opendaylight.controller.cluster.raft.messages.AddServer; import org.opendaylight.controller.cluster.raft.messages.AddServerReply; -import org.opendaylight.controller.cluster.raft.messages.FollowerCatchUpTimeout; +import org.opendaylight.controller.cluster.raft.messages.ChangeServersVotingStatus; +import org.opendaylight.controller.cluster.raft.messages.RemoveServer; +import org.opendaylight.controller.cluster.raft.messages.RemoveServerReply; +import org.opendaylight.controller.cluster.raft.messages.ServerChangeReply; import org.opendaylight.controller.cluster.raft.messages.ServerChangeStatus; +import org.opendaylight.controller.cluster.raft.messages.ServerRemoved; import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply; +import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload; +import org.opendaylight.controller.cluster.raft.persisted.ServerInfo; 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; @@ -37,69 +50,107 @@ import scala.concurrent.duration.FiniteDuration; */ class RaftActorServerConfigurationSupport { private static final Logger LOG = LoggerFactory.getLogger(RaftActorServerConfigurationSupport.class); - private final RaftActorContext context; - // client follower queue - private final Queue followerInfoQueue = new LinkedList(); - // timeout handle - private Cancellable followerTimeout = null; - RaftActorServerConfigurationSupport(RaftActorContext context) { - this.context = context; + @SuppressWarnings("checkstyle:MemberName") + private final OperationState IDLE = new Idle(); + + private final RaftActor raftActor; + + private final RaftActorContext raftContext; + + private final Queue> pendingOperationsQueue = new ArrayDeque<>(); + + private OperationState currentOperationState = IDLE; + + RaftActorServerConfigurationSupport(final RaftActor raftActor) { + this.raftActor = raftActor; + this.raftContext = raftActor.getRaftActorContext(); } - boolean handleMessage(Object message, RaftActor raftActor, ActorRef sender) { - if(message instanceof AddServer) { - onAddServer((AddServer)message, raftActor, sender); + boolean handleMessage(final Object message, final ActorRef sender) { + if (message instanceof AddServer) { + onAddServer((AddServer) message, sender); + return true; + } else if (message instanceof RemoveServer) { + onRemoveServer((RemoveServer) message, sender); return true; - } else if (message instanceof FollowerCatchUpTimeout){ - FollowerCatchUpTimeout followerTimeout = (FollowerCatchUpTimeout)message; - // abort follower catchup on timeout - onFollowerCatchupTimeout(raftActor, sender, followerTimeout.getNewServerId()); + } else if (message instanceof ChangeServersVotingStatus) { + onChangeServersVotingStatus((ChangeServersVotingStatus) message, sender); return true; - } else if (message instanceof UnInitializedFollowerSnapshotReply){ - // snapshot installation is successful - onUnInitializedFollowerSnapshotReply((UnInitializedFollowerSnapshotReply)message, raftActor,sender); + } else if (message instanceof ServerOperationTimeout) { + currentOperationState.onServerOperationTimeout((ServerOperationTimeout) message); return true; - } else if(message instanceof ApplyState) { - return onApplyState((ApplyState) message, raftActor); + } else if (message instanceof UnInitializedFollowerSnapshotReply) { + currentOperationState.onUnInitializedFollowerSnapshotReply((UnInitializedFollowerSnapshotReply) message); + return true; + } else if (message instanceof ApplyState) { + return onApplyState((ApplyState) message); + } else if (message instanceof SnapshotComplete) { + currentOperationState.onSnapshotComplete(); + return false; } else { return false; } } - private boolean onApplyState(ApplyState applyState, RaftActor raftActor) { - Payload data = applyState.getReplicatedLogEntry().getData(); - if(data instanceof ServerConfigurationPayload) { - CatchupFollowerInfo followerInfo = followerInfoQueue.peek(); - if(followerInfo != null && followerInfo.getContextId().equals(applyState.getIdentifier())) { - LOG.info("{} has been successfully replicated to a majority of followers", data); - - // respond ok to follower - respondToClient(raftActor, ServerChangeStatus.OK); - } + void onNewLeader(final String leaderId) { + currentOperationState.onNewLeader(leaderId); + } - return true; + private void onChangeServersVotingStatus(final ChangeServersVotingStatus message, final ActorRef sender) { + LOG.debug("{}: onChangeServersVotingStatus: {}, state: {}", raftContext.getId(), message, + currentOperationState); + + // 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)); } - - return false; } - private void onAddServer(AddServer addServer, RaftActor raftActor, ActorRef sender) { - LOG.debug("{}: onAddServer: {}", context.getId(), addServer); - if(noLeaderOrForwardedToLeader(addServer, raftActor, sender)) { - return; + private void onRemoveServer(final RemoveServer removeServer, final ActorRef sender) { + LOG.debug("{}: onRemoveServer: {}, state: {}", raftContext.getId(), removeServer, currentOperationState); + boolean isSelf = removeServer.getServerId().equals(raftContext.getId()); + if (isSelf && !raftContext.hasFollowers()) { + sender.tell(new RemoveServerReply(ServerChangeStatus.NOT_SUPPORTED, raftActor.getLeaderId()), + raftActor.getSelf()); + } else if (!isSelf && !raftContext.getPeerIds().contains(removeServer.getServerId())) { + sender.tell(new RemoveServerReply(ServerChangeStatus.DOES_NOT_EXIST, raftActor.getLeaderId()), + raftActor.getSelf()); + } else { + String serverAddress = isSelf ? raftActor.self().path().toString() : + raftContext.getPeerAddress(removeServer.getServerId()); + onNewOperation(new RemoveServerContext(removeServer, serverAddress, sender)); } + } - CatchupFollowerInfo followerInfo = new CatchupFollowerInfo(addServer,sender); - boolean process = followerInfoQueue.isEmpty(); - followerInfoQueue.add(followerInfo); - if(process) { - processAddServer(raftActor); + private boolean onApplyState(final ApplyState applyState) { + Payload data = applyState.getReplicatedLogEntry().getData(); + if (data instanceof ServerConfigurationPayload) { + currentOperationState.onApplyState(applyState); + return true; } + + return false; } /** - * The algorithm for AddServer is as follows: + * Add a server. The algorithm for AddServer is as follows: *
    *
  • Add the new server as a peer.
    • *
  • Add the new follower to the leader.
    • @@ -119,132 +170,722 @@ class RaftActorServerConfigurationSupport { *
  • Respond to caller with TIMEOUT.
    • *
*/ - private void processAddServer(RaftActor raftActor){ - LOG.debug("{}: In processAddServer", context.getId()); - - AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior(); - CatchupFollowerInfo followerInfo = followerInfoQueue.peek(); - AddServer addSrv = followerInfo.getAddServer(); - context.addToPeers(addSrv.getNewServerId(), addSrv.getNewServerAddress()); - - // if voting member - initialize to VOTING_NOT_INITIALIZED - FollowerState initialState = addSrv.isVotingMember() ? FollowerState.VOTING_NOT_INITIALIZED : - FollowerState.NON_VOTING; - leader.addFollower(addSrv.getNewServerId(), initialState); - - if(initialState == FollowerState.VOTING_NOT_INITIALIZED){ - LOG.debug("Leader sending initiate capture snapshot to follower : {}", addSrv.getNewServerId()); - leader.initiateCaptureSnapshot(addSrv.getNewServerId()); - // schedule the catchup timeout timer - followerTimeout = context.getActorSystem().scheduler() - .scheduleOnce(new FiniteDuration(((context.getConfigParams().getElectionTimeOutInterval().toMillis()) * 2), - TimeUnit.MILLISECONDS), - context.getActor(), new FollowerCatchUpTimeout(addSrv.getNewServerId()), - context.getActorSystem().dispatcher(), context.getActor()); + private void onAddServer(final AddServer addServer, final ActorRef sender) { + LOG.debug("{}: onAddServer: {}, state: {}", raftContext.getId(), addServer, currentOperationState); + + onNewOperation(new AddServerContext(addServer, sender)); + } + + private void onNewOperation(final ServerOperationContext operationContext) { + if (raftActor.isLeader()) { + currentOperationState.onNewOperation(operationContext); } else { - LOG.debug("Directly persisting the new server configuration : {}", addSrv.getNewServerId()); - persistNewServerConfiguration(raftActor, followerInfo); + ActorSelection leader = raftActor.getLeader(); + if (leader != null) { + LOG.debug("{}: Not leader - forwarding to leader {}", raftContext.getId(), leader); + leader.tell(operationContext.getOperation(), operationContext.getClientRequestor()); + } else { + LOG.debug("{}: No leader - returning NO_LEADER reply", raftContext.getId()); + operationContext.getClientRequestor().tell(operationContext.newReply( + ServerChangeStatus.NO_LEADER, null), raftActor.self()); + } } } - private boolean noLeaderOrForwardedToLeader(Object message, RaftActor raftActor, ActorRef sender) { - if (raftActor.isLeader()) { - return false; + /** + * 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(final 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); } - ActorSelection leader = raftActor.getLeader(); - if (leader != null) { - LOG.debug("Not leader - forwarding to leader {}", leader); - leader.forward(message, raftActor.getContext()); - } else { - LOG.debug("No leader - returning NO_LEADER AddServerReply"); - sender.tell(new AddServerReply(ServerChangeStatus.NO_LEADER, null), raftActor.self()); + void onServerOperationTimeout(final ServerOperationTimeout timeout) { + LOG.debug("onServerOperationTimeout should not be called in state {}", this); + } + + void onUnInitializedFollowerSnapshotReply(final UnInitializedFollowerSnapshotReply reply) { + LOG.debug("onUnInitializedFollowerSnapshotReply was called in state {}", this); + } + + void onApplyState(final ApplyState applyState) { + LOG.debug("onApplyState was called in state {}", this); + } + + void onSnapshotComplete() { + + } + + void onNewLeader(final String newLeader) { + } + + protected void persistNewServerConfiguration(final 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); } - return true; + protected void operationComplete(final ServerOperationContext operationContext, + final @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(final ServerOperationContext operationContext, final ServerChangeStatus status) { + LOG.debug("{}: Returning {} for operation {}", raftContext.getId(), status, + operationContext.getOperation()); + + operationContext.getClientRequestor().tell(operationContext.newReply(status, raftActor.getLeaderId()), + raftActor.self()); + } + + Cancellable newTimer(final Object message) { + return newTimer(raftContext.getConfigParams().getElectionTimeOutInterval().$times(2), message); + } + + Cancellable newTimer(final FiniteDuration timeout, final 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(final ServerOperationContext operationContext) { + operationContext.newInitialOperationState(RaftActorServerConfigurationSupport.this).initiate(); + } + + @Override + public void onApplyState(final 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(final ServerOperationContext operationContext, final Cancellable timer) { + this.operationContext = operationContext; + this.timer = timer; + } + + @Override + public void onApplyState(final 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(final 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(final ServerOperationContext newOperationContext) { + if (timedOut) { + sendReply(newOperationContext, ServerChangeStatus.PRIOR_REQUEST_CONSENSUS_TIMEOUT); + } else { + super.onNewOperation(newOperationContext); + } + } } - private void onUnInitializedFollowerSnapshotReply(UnInitializedFollowerSnapshotReply reply, - RaftActor raftActor, ActorRef sender){ - CatchupFollowerInfo followerInfo = followerInfoQueue.peek(); - // Sanity check - it's possible we get a reply after it timed out. - if(followerInfo == null) { - return; + /** + * Abstract base class for an AddServer operation state. + */ + private abstract class AddServerState extends OperationState { + private final AddServerContext addServerContext; + + AddServerState(final AddServerContext addServerContext) { + this.addServerContext = addServerContext; + } + + AddServerContext getAddServerContext() { + return addServerContext; + } + + Cancellable newInstallSnapshotTimer() { + return newTimer(new ServerOperationTimeout(addServerContext.getOperation().getNewServerId())); } - String followerId = reply.getFollowerId(); - AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior(); - FollowerLogInformation followerLogInformation = leader.getFollower(followerId); - stopFollowerTimer(); - followerLogInformation.setFollowerState(FollowerState.VOTING); - leader.updateMinReplicaCountAndMinIsolatedLeaderPeerCount(); + void handleInstallSnapshotTimeout(final 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); + } - persistNewServerConfiguration(raftActor, followerInfo); } - private void persistNewServerConfiguration(RaftActor raftActor, CatchupFollowerInfo followerInfo){ - List cNew = new ArrayList(context.getPeerAddresses().keySet()); - cNew.add(context.getId()); + /** + * 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(final 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()); - LOG.debug("New server configuration : {}", cNew.toString()); + persistNewServerConfiguration(getAddServerContext()); + } + } + } - ServerConfigurationPayload servPayload = new ServerConfigurationPayload(cNew, Collections.emptyList()); + /** + * 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; - raftActor.persistData(followerInfo.getClientRequestor(), followerInfo.getContextId(), servPayload); - } + InstallingSnapshot(final AddServerContext addServerContext, final Cancellable installSnapshotTimer) { + super(addServerContext); + this.installSnapshotTimer = requireNonNull(installSnapshotTimer); + } + + @Override + public void onServerOperationTimeout(final ServerOperationTimeout timeout) { + handleInstallSnapshotTimeout(timeout); - private void stopFollowerTimer() { - if (followerTimeout != null && !followerTimeout.isCancelled()) { - followerTimeout.cancel(); + LOG.warn("{}: Timeout occured for new server {} while installing snapshot", raftContext.getId(), + timeout.getLoggingContext()); } - } - private void onFollowerCatchupTimeout(RaftActor raftActor, ActorRef sender, String serverId){ - LOG.debug("onFollowerCatchupTimeout: {}", serverId); - AbstractLeader leader = (AbstractLeader) raftActor.getCurrentBehavior(); - // cleanup - context.removePeer(serverId); - leader.removeFollower(serverId); - LOG.warn("Timeout occured for new server {} while installing snapshot", serverId); - respondToClient(raftActor,ServerChangeStatus.TIMEOUT); - } + @Override + public void onUnInitializedFollowerSnapshotReply(final UnInitializedFollowerSnapshotReply reply) { + LOG.debug("{}: onUnInitializedFollowerSnapshotReply: {}", raftContext.getId(), reply); - private void respondToClient(RaftActor raftActor, ServerChangeStatus result){ - // remove the entry from the queue - CatchupFollowerInfo fInfo = followerInfoQueue.remove(); + String followerId = reply.getFollowerId(); - // get the sender - ActorRef toClient = fInfo.getClientRequestor(); + // 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(); - toClient.tell(new AddServerReply(result, raftActor.getLeaderId()), raftActor.self()); - LOG.debug("Response returned is {} for server {} ", result, fInfo.getAddServer().getNewServerId()); - if(!followerInfoQueue.isEmpty()){ - processAddServer(raftActor); + 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; - // maintain sender actorRef - private static class CatchupFollowerInfo { - private final AddServer addServer; + WaitingForPriorSnapshotComplete(final AddServerContext addServerContext, final Cancellable snapshotTimer) { + super(addServerContext); + this.snapshotTimer = requireNonNull(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(final 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 { + private static final long serialVersionUID = 1L; + + ServerOperationContextIdentifier() { + super(UUID.randomUUID()); + } + } + + /** + * Stores context information for a server operation. + * + * @param the operation type + */ + private abstract static class ServerOperationContext { + private final T operation; private final ActorRef clientRequestor; - private final String contextId; + private final Identifier contextId; - CatchupFollowerInfo(AddServer addSrv, ActorRef cliReq){ - addServer = addSrv; - clientRequestor = cliReq; - contextId = UUID.randomUUID().toString(); + ServerOperationContext(final T operation, final ActorRef clientRequestor) { + this.operation = operation; + this.clientRequestor = clientRequestor; + contextId = new ServerOperationContextIdentifier(); } - String getContextId() { + Identifier getContextId() { return contextId; } - AddServer getAddServer(){ - return addServer; + T getOperation() { + return operation; } - ActorRef getClientRequestor(){ + ActorRef getClientRequestor() { return clientRequestor; } + + void operationComplete(final RaftActor raftActor, final boolean succeeded) { + } + + boolean includeSelfInNewConfiguration(final 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 { + AddServerContext(final AddServer addServer, final ActorRef clientRequestor) { + super(addServer, clientRequestor); + } + + @Override + Object newReply(final ServerChangeStatus status, final String leaderId) { + return new AddServerReply(status, leaderId); + } + + @Override + InitialOperationState newInitialOperationState(final RaftActorServerConfigurationSupport support) { + return support.new InitialAddServerState(this); + } + + @Override + String getLoggingContext() { + return getOperation().getNewServerId(); + } + } + + private abstract class RemoveServerState extends OperationState { + private final RemoveServerContext removeServerContext; + + protected RemoveServerState(final RemoveServerContext removeServerContext) { + this.removeServerContext = requireNonNull(removeServerContext); + + } + + public RemoveServerContext getRemoveServerContext() { + return removeServerContext; + } + } + + private final class InitialRemoveServerState extends RemoveServerState implements InitialOperationState { + + protected InitialRemoveServerState(final RemoveServerContext removeServerContext) { + super(removeServerContext); + } + + @Override + public void initiate() { + String serverId = getRemoveServerContext().getOperation().getServerId(); + raftContext.removePeer(serverId); + AbstractLeader leader = (AbstractLeader)raftActor.getCurrentBehavior(); + leader.removeFollower(serverId); + leader.updateMinReplicaCount(); + + persistNewServerConfiguration(getRemoveServerContext()); + } + } + + private static class RemoveServerContext extends ServerOperationContext { + private final String peerAddress; + + RemoveServerContext(final RemoveServer operation, final String peerAddress, final ActorRef clientRequestor) { + super(operation, clientRequestor); + this.peerAddress = peerAddress; + } + + @Override + Object newReply(final ServerChangeStatus status, final String leaderId) { + return new RemoveServerReply(status, leaderId); + } + + @Override + InitialOperationState newInitialOperationState(final RaftActorServerConfigurationSupport support) { + return support.new InitialRemoveServerState(this); + } + + @Override + void operationComplete(final RaftActor raftActor, final boolean succeeded) { + if (peerAddress != null) { + raftActor.context().actorSelection(peerAddress).tell( + new ServerRemoved(getOperation().getServerId()), raftActor.getSelf()); + } + } + + @Override + boolean includeSelfInNewConfiguration(final RaftActor raftActor) { + return !getOperation().getServerId().equals(raftActor.getId()); + } + + @Override + String getLoggingContext() { + return getOperation().getServerId(); + } + } + + private static class ChangeServersVotingStatusContext extends ServerOperationContext { + private final boolean tryToElectLeader; + + ChangeServersVotingStatusContext(final ChangeServersVotingStatus convertMessage, final ActorRef clientRequestor, + final boolean tryToElectLeader) { + super(convertMessage, clientRequestor); + this.tryToElectLeader = tryToElectLeader; + } + + @Override + InitialOperationState newInitialOperationState(final RaftActorServerConfigurationSupport support) { + return support.new ChangeServersVotingStatusState(this, tryToElectLeader); + } + + @Override + Object newReply(final ServerChangeStatus status, final String leaderId) { + return new ServerChangeReply(status, leaderId); + } + + @Override + void operationComplete(final RaftActor raftActor, final boolean succeeded) { + // If this leader changed to non-voting we need to step down as leader so we'll try to transfer + // leadership. + boolean localServerChangedToNonVoting = Boolean.FALSE.equals(getOperation() + .getServerVotingStatusMap().get(raftActor.getRaftActorContext().getId())); + if (succeeded && localServerChangedToNonVoting) { + LOG.debug("Leader changed to non-voting - trying leadership transfer"); + raftActor.becomeNonVoting(); + } else if (raftActor.isLeader()) { + raftActor.onVotingStateChangeComplete(); + } + } + + @Override + String getLoggingContext() { + return getOperation().toString(); + } + } + + private class ChangeServersVotingStatusState extends OperationState implements InitialOperationState { + private final ChangeServersVotingStatusContext changeVotingStatusContext; + private final boolean tryToElectLeader; + + ChangeServersVotingStatusState(final ChangeServersVotingStatusContext changeVotingStatusContext, + final 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(TimeoutNow.INSTANCE, raftContext.getActor()); + + currentOperationState = new WaitingForLeaderElected(changeVotingStatusContext, previousServerConfig); + } + + private boolean updateLocalPeerInfo() { + List 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 newServerInfoList() { + Map serverVotingStatusMap = changeVotingStatusContext.getOperation() + .getServerVotingStatusMap(); + List newServerInfoList = new ArrayList<>(); + for (String peerId: raftContext.getPeerIds()) { + newServerInfoList.add(new ServerInfo(peerId, serverVotingStatusMap.containsKey(peerId) + ? serverVotingStatusMap.get(peerId) : raftContext.getPeerInfo(peerId).isVoting())); + } + + newServerInfoList.add(new ServerInfo(raftContext.getId(), serverVotingStatusMap.containsKey( + raftContext.getId()) ? serverVotingStatusMap.get(raftContext.getId()) + : raftContext.isVotingMember())); + + return newServerInfoList; + } + } + + private class WaitingForLeaderElected extends OperationState { + private final ServerConfigurationPayload previousServerConfig; + private final ChangeServersVotingStatusContext operationContext; + private final Cancellable timer; + + WaitingForLeaderElected(final ChangeServersVotingStatusContext operationContext, + final ServerConfigurationPayload previousServerConfig) { + this.operationContext = operationContext; + this.previousServerConfig = previousServerConfig; + + timer = newTimer(raftContext.getConfigParams().getElectionTimeOutInterval(), + new ServerOperationTimeout(operationContext.getLoggingContext())); + } + + @Override + void onNewLeader(final String newLeader) { + if (newLeader == null) { + return; + } + + 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); + + changeToIdleState(); + RaftActorServerConfigurationSupport.this.onNewOperation(operationContext); + } + } + + @Override + void onServerOperationTimeout(final 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 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 serverVotingStatusMap = operationContext.getOperation().getServerVotingStatusMap(); + ActorSelection forwardToPeerActor = null; + for (Map.Entry 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); + } + } + } + + static class ServerOperationTimeout { + private final String loggingContext; + + ServerOperationTimeout(final String loggingContext) { + this.loggingContext = requireNonNull(loggingContext, "loggingContext should not be null"); + } + + String getLoggingContext() { + return loggingContext; + } } }