CSR: Small: Client-Centric Consistency

Cornell University, Ithaca NY

Investigators

Abstract

The isolation guarantees offered by a storage system regulate how multiple transactions executing concurrently read and write data correctly. Similarly, when data is replicated and updated at geographically-distinct locations (as do Facebook and Twitter), consistency guarantees specify which data is visible to clients that access different replicas. This project will develop a new programmer-centric model for expressing rigorously, in a single framework, the isolation and consistency guarantees offered by geo-replicated storage systems. It will then leverage the new model to seek more efficient ways to implement these guarantees. The project will proceed along three thrusts. First, it will seek to express isolation and consistency guarantees as properties of states that applications can observe, instead of relying on invisible histories of low-level read/write operations. Second, it will seek to overcome the current scalability bottlenecks that arise with slow or faulty servers, which have, to date, stifled industry's interest in deploying causally-consistent geo-replicated data stores. Third, it will leverage its unified treatment of consistency and isolation to support, for the first time, generic read/write transactions on top of a scalable consistent data store. The project directly addresses the growing interest in both research and industry towards large-scale, transactional distributed systems. By confronting key obstacles to scalability, it will contribute to bring to industrial fruition over two decades of academic research on causal consistency. Further, by establishing a common language to discuss consistency and isolation, it will enable more fruitful exchanges between the systems and database research communities, who, because of their primary focus on, respectively, consistency and isolation, have often been talking past each other. The project, with its tight integration of both theoretical and experimental research, is ideally suited to introducing graduate and undergraduate students to research, and will offer a new approach to teaching consistency and isolation in graduate and undergraduate courses in distributed computing. The project outcomes (papers published, presentations, and a repository of the software and workloads used to demonstrate the results of the project) will be made available for download at the address http://www.cs.utexas.edu/users/lorenzo/projects/CCC/, where they will be kept for five years.