METRIC: A Middleware for Entry Transactional Database Clustering at the Edge

Abstract

A geo-distributed database for edge architectures spanning thousands of sites needs to assure efficient local updates while replicating sufficient state across sites to enable global management and support mobility, failover etc. To address this requirement, a new paradigm for database clustering that achieves a better balance than existing solutions between performance and strength of semantics called entry transactionality is introduced. Inspired by entry consistency in shared memory systems, entry transactionality guarantees that only a client that owns a range of keys in the database has a sequentially consistent value of the keys and can perform local and, hence, efficient transactions across these keys. Important use cases enabled by entry transactionality such as federated controllers and state management for edge applications are identified. The semantics of entry transactionality incorporating the complex failure modes in geo-distributed services are defined, and the difficult challenges in realizing these semantics are outlined. Then, a novel Middleware for Entry Transactional Clustering (METRIC) that combines existing SQL databases with an underlying geo-distributed entry consistent store to realize entry transactionality is described. This paper provides initial findings from an on-going effort.