Brief Announcement: Multicoordinated Paxos∗

In the consensus problem, processes must agree on a single value, given a set of proposals. It is used, for example, to implement statemachine replication [4], in which failure-independent processors are replicated to implement a reliable service. In this approach, the states of the replicas are consistently changed by applying deterministic commands from a sequence, each one agreed upon using a consensus instance. Paxos [5] is an efficient and fault-tolerant consensus protocol originally intended for state-machine replication. It relies on a leader process to receive proposals from proposer processes (e.g., clients) and forward them to be agreed by the acceptors, which then notify the learners (e.g., replicas). Even though Paxos provides very good performance, getting commands delivered by replicas in three communication steps, normal execution depends on the availability of the current leader. If the leader fails (or seems to have failed) a new leader must be elected, and this new leader has to synchronize with a quorum of acceptors before resuming normal execution. These actions take time and may introduce some temporary unavailability to the system. While this unavailability may not be significant when running a single instance of Paxos, it may cause a considerable slowdown on real systems, which use infinitely many instances of them [1]. Multicoordinated Paxos is an improved version of Paxos in which the failure of the leader does not slow down the protocol and the acceptance of new proposals. Our improvements neither increase Paxos latency nor require more acceptors to be contacted. In Multicoordinated Paxos, proposals are sent not to the leader, but to a set of coordinator processes, which act like the leader, and forward proposals to the acceptors. Acceptors, however, will only acknowledge values that have been sent by a quorum of coordinators. Fast Paxos [7] is another improved variant of Paxos. In good runs of Fast Paxos, proposers send their proposals directly to the acceptors, reducing the minimum time to get a command learned to two communication steps. Since the leader can be bypassed, its unavailability is much less disruptive to the system, as in Multi-