Brief Announcement: MUSIC: Multi-Site Entry Consistencyfor Geo-Distributed Services

Abstract

Geo-distributed services that are executed across multiple sites at a global scale are increasingly prevalent, and are at the core of the control plane tasked with managing the Virtual Network Functions (VNFs) of next-generation network infrastructure. A key building block of any replicated service of this type is logically shared state, an abstraction that is often implemented by having multiple copies of each data element and using one of any number of distributed protocols to enforce consistency with guarantees ranging from eventual to causal to sequential consistency. These solutions fail to meet the real needs of geo-distributed services found in a VNF control plane (and elsewhere), which are to provide guarantees strong enough to be useful, yet still efficiently implementable in a global network with larger latencies and a wider range of failure modes. Here, we propose a new solution to state consistency based on extending entry consistency, a paradigm originally developed for failure-free multi-processor systems, for use in geo-distributed services with failures. We outline the challenges associated with state consistency in such services, and describe a data-store called MUSIC (MUlti-SIte entry Consistency) that combines an eventually consistent key-value store with locking primitives to implement entry-consistent semantics.