Class KetchLeader


  • public abstract class KetchLeader
    extends Object
    A leader managing consensus across remote followers.

    A leader instance starts up in KetchLeader.State.CANDIDATE and tries to begin a new term by sending an ElectionRound to all replicas. Its term starts if a majority of replicas have accepted this leader instance for the term.

    Once elected by a majority the instance enters KetchLeader.State.LEADER and runs proposals offered to queueProposal(Proposal). This continues until the leader is timed out for inactivity, or is deposed by a competing leader gaining its own majority.

    Once timed out or deposed this KetchLeader instance should be discarded, and a new instance takes over.

    Each leader instance coordinates a group of KetchReplicas. Replica instances are owned by the leader instance and must be discarded when the leader is discarded.

    In Ketch all push requests are issued through the leader. The steps are as follows (see KetchPreReceive for an example):

    The leader gains consensus by first pushing the needed objects and a RefTree representing the desired target repository state to the refs/txn/accepted branch on each of the replicas. Once a majority has succeeded, the leader commits the state by either pushing the refs/txn/accepted value to refs/txn/committed (for Ketch-aware replicas) or by pushing updates to refs/heads/master, etc. for stock Git replicas.

    Internally, the actual transport to replicas is performed on background threads via the KetchSystem's executor service. For performance, the KetchLeader, KetchReplica and Proposal objects share some state, and may invoke each other's methods on different threads. This access is protected by the leader's lock object. Care must be taken to prevent concurrent access by correctly obtaining the leader's lock.

    • Constructor Detail

      • KetchLeader

        protected KetchLeader​(KetchSystem system)
        Construct a leader for a Ketch instance.
        Parameters:
        system - Ketch system configuration the leader must adhere to.
    • Method Detail

      • setReplicas

        public void setReplicas​(Collection<KetchReplica> replicas)
        Configure the replicas used by this Ketch instance.

        Replicas should be configured once at creation before any proposals are executed. Once elections happen, reconfiguration is a complicated concept that is not currently supported.

        Parameters:
        replicas - members participating with the same repository.
      • openRepository

        protected abstract Repository openRepository()
                                              throws IOException
        Get an instance of the repository for use by a leader thread.

        The caller will close the repository.

        Returns:
        opened repository for use by the leader thread.
        Throws:
        IOException - cannot reopen the repository for the leader.
      • queueProposal

        public void queueProposal​(Proposal proposal)
                           throws InterruptedException,
                                  IOException
        Queue a reference update proposal for consensus.

        This method does not wait for consensus to be reached. The proposal is checked to look for risks of conflicts, and then submitted into the queue for distribution as soon as possible.

        Callers must use Proposal.await() to see if the proposal is done.

        Parameters:
        proposal - the proposed reference updates to queue for consideration. Once execution is complete the individual reference result fields will be populated with the outcome.
        Throws:
        InterruptedException - current thread was interrupted. The proposal may have been aborted if it was not yet queued for execution.
        IOException - unrecoverable error preventing proposals from being attempted by this leader.
      • snapshot

        public LeaderSnapshot snapshot()
        Snapshot this leader
        Returns:
        snapshot of this leader
      • shutdown

        public void shutdown()
        Gracefully shutdown this leader and cancel outstanding operations.