Package org.opendaylight.controller.cluster.raft

Class RaftActor

  • All Implemented Interfaces:
    Actor, Stash, StashFactory, StashSupport, UnrestrictedStash, RequiresMessageQueue<DequeBasedMessageQueueSemantics>, AbstractPersistentActorLike, Eventsourced, PersistenceIdentity, PersistenceRecovery, PersistenceStash, Snapshotter, ExecuteInSelfActor
    Direct Known Subclasses:
    Shard
    public abstract class RaftActor
extends AbstractUntypedPersistentActor
    RaftActor encapsulates a state machine that needs to be kept synchronized in a cluster. It implements the RAFT algorithm as described in the paper In Search of an Understandable Consensus Algorithm

    RaftActor has 3 states and each state has a certain behavior associated with it. A Raft actor can behave as,

    • A Leader
    • A Follower (or)
    • A Candidate

    A RaftActor MUST be a Leader in order to accept requests from clients to change the state of it's encapsulated state machine. Once a RaftActor becomes a Leader it is also responsible for ensuring that all followers ultimately have the same log and therefore the same state machine as itself.

    The current behavior of a RaftActor determines how election for leadership is initiated and how peer RaftActors react to request for votes.

    Each RaftActor also needs to know the current election term. It uses this information for a couple of things. One is to simply figure out who it voted for in the last election. Another is to figure out if the message it received to update it's state is stale.

    The RaftActor uses akka-persistence to store it's replicated log. Furthermore through it's behaviors a Raft Actor determines

    • when a log entry should be persisted
    • when a log entry should be applied to the state machine (and)
    • when a snapshot should be saved

    • Method Detail

      • newRaftActorRecoverySupport

        protected org.opendaylight.controller.cluster.raft.RaftActorRecoverySupport newRaftActorRecoverySupport()

      • changeCurrentBehavior

        protected void changeCurrentBehavior​(RaftActorBehavior newBehavior)

      • handleNonRaftCommand

        protected void handleNonRaftCommand​(Object message)
        Method exposed for subclasses to plug-in their logic. This method is invoked by handleCommand(Object) for messages which are not handled by this class. Subclasses overriding this class should fall back to this implementation for messages which they do not handle
        Parameters:
        message - Incoming command message

      • persistData

        protected final void persistData​(ActorRef clientActor,
                                 Identifier identifier,
                                 Payload data,
                                 boolean batchHint)
        Persists the given Payload in the journal and replicates to any followers. After successful completion, applyState(ActorRef, Identifier, Object) is notified.
        Parameters:
        clientActor - optional ActorRef that is provided via the applyState callback
        identifier - the payload identifier
        data - the payload data to persist
        batchHint - if true, an attempt is made to delay immediate replication and batch the payload with subsequent payloads for efficiency. Otherwise the payload is immediately replicated.

      • getId

        protected String getId()

      • isLeader

        protected boolean isLeader()
        Derived actors can call the isLeader method to check if the current RaftActor is the Leader or not.
        Returns:
        true it this RaftActor is a Leader false otherwise

      • isLeaderActive

        protected final boolean isLeaderActive()

      • isLeadershipTransferInProgress

        protected boolean isLeadershipTransferInProgress()

      • getLeader

        public ActorSelection getLeader()
        Derived actor can call getLeader if they need a reference to the Leader. This would be useful for example in forwarding a request to an actor which is the leader
        Returns:
        A reference to the leader if known, null otherwise

      • getLeaderId

        protected final String getLeaderId()
        Returns the id of the current leader.
        Returns:
        the current leader's id

      • getRaftState

        protected final RaftState getRaftState()

      • getCurrentTerm

        protected Long getCurrentTerm()

      • updateConfigParams

        protected void updateConfigParams​(ConfigParams configParams)

      • setPersistence

        protected void setPersistence​(boolean persistent)

      • setPeerAddress

        protected void setPeerAddress​(String peerId,
                              String peerAddress)
        setPeerAddress sets the address of a known peer at a later time.

        This is to account for situations where a we know that a peer exists but we do not know an address up-front. This may also be used in situations where a known peer starts off in a different location and we need to change it's address

        Note that if the peerId does not match the list of peers passed to this actor during construction an IllegalStateException will be thrown.

      • applyState

        protected abstract void applyState​(ActorRef clientActor,
                                   Identifier identifier,
                                   Object data)
        The applyState method will be called by the RaftActor when some data needs to be applied to the actor's state.
        Parameters:
        clientActor - A reference to the client who sent this message. This is the same reference that was passed to persistData by the derived actor. clientActor may be null when the RaftActor is behaving as a follower or during recovery.
        identifier - The identifier of the persisted data. This is also the same identifier that was passed to persistData by the derived actor. identifier may be null when the RaftActor is behaving as a follower or during recovery
        data - A piece of data that was persisted by the persistData call. This should NEVER be null.

      • getRaftActorRecoveryCohort

        protected abstract @NonNull RaftActorRecoveryCohort getRaftActorRecoveryCohort()
        Returns the RaftActorRecoveryCohort to participate in persistence recovery.

      • onRecoveryComplete

        protected abstract void onRecoveryComplete()
        This method is called when recovery is complete.

      • getRaftActorSnapshotCohort

        protected abstract @NonNull RaftActorSnapshotCohort getRaftActorSnapshotCohort()
        Returns the RaftActorSnapshotCohort to participate in snapshot captures.

      • onStateChanged

        protected abstract void onStateChanged()
        This method will be called by the RaftActor when the state of the RaftActor changes. The derived actor can then use methods like isLeader or getLeader to do something useful

      • getRoleChangeNotifier

        protected abstract Optional<ActorRef> getRoleChangeNotifier()
        Notifier Actor for this RaftActor to notify when a role change happens.
        Returns:
        ActorRef - ActorRef of the notifier or Optional.absent if none.

      • onVotingStateChangeComplete

        protected void onVotingStateChangeComplete()
        This method is called on the leader when a voting change operation completes.

      • pauseLeader

        protected void pauseLeader​(Runnable operation)
        This method is called prior to operations such as leadership transfer and actor shutdown when the leader must pause or stop its duties. This method allows derived classes to gracefully pause or finish current work prior to performing the operation. On completion of any work, the run method must be called on the given Runnable to proceed with the given operation. Important: the run method must be called on this actor's thread dispatcher as as it modifies internal state.

        The default implementation immediately runs the operation.

        Parameters:
        operation - the operation to run

      • unpauseLeader

        protected void unpauseLeader()
        This method is invoked when the actions hooked to the leader becoming paused failed to execute and the leader should resume normal operations.

        Note this method can be invoked even before the operation supplied to pauseLeader(Runnable) is invoked.

      • onLeaderChanged

        protected void onLeaderChanged​(String oldLeader,
                               String newLeader)

      • hasFollowers

        protected boolean hasFollowers()