Amar Phanishayee

Chaining For Flexible And High-Performance Key-Value Systems Degree Type: Ph.D. in Computer Science
Advisor(s): David Andersen
Graduated: December 2012

Abstract:

Distributed key-value (KV) systems are a critical part of the infrastructure at many large sites such as Amazon, Facebook, Google, and Twitter. The first research question this dissertation addresses is: How should we design a cluster-based key-value store that is fault tolerant, achieves high performance and availability, and offers strong data consistency? We present a new replication protocol, Ouroboros, which extends chain-based replication to allow fast, non-blocking node additions to any part of the replica chain, and guarantees provably strong data consistency. We use Ouroboros to implement a distributed key-value system, FAWN-KV, designed with the goal of supporting the three key properties of fault tolerance, high performance, and generality. We present a formal proof of correctness of Ouroboros, and evaluate FAWN-KV on clusters with Flash storage.

FAWN-KV is, still, only a specific KV solution that offers strong data consistency and is optimized for clusters that have storage devices with slow random writes. The current diversity in hardware and application requirements have resulted in a plethora of KV systems today, with no one system meeting the needs of all applications. The second, and final, research question this dissertation addresses is therefore: Is it possible for a KV architecture to be easily configured to support many points along the KV system design continuum? We present a generalization of chain-based replication, Ouroboros+, which extends Ouroboros to effectively support a wide range of application requirements by (a) selecting from different update protocols between replicas, and, (b) selecting a query node in a replica chain. We describe Flex-KV, which uses Ouroboros+ with different datastores that expose a common storage interface to form heterogeneous replica chains. Flex-KV can support DRAM, Flash, and disk-based storage; can act as an unreliable cache or a durable store; and can offer strong or weak data consistency. e value of such a system goes beyond ease-of-use: We enable new choices for system designs that offer some applications a better balance of performance and cost than was previously available

Thesis Committee:
David G. Andersen (Chair)
Garth A. Gibson
Srinivasan Seshan
Timothy Roscoe (ETH Zurich)

Jeannette Wing, Head, Computer Science Department
Randy Bryant, Dean, School of Computer Science

Keywords:
Key-Value Systems, Replication Protocols, Data Consistency, Flash