Resilient Composition for Fault-Tolerant Cloud Applications

Abstract

As the cloud evolves in capability, it has also become increasingly complex and difficult to program. New abstractions are necessary to ensure next-generation cloud applications are correct, simple, and efficient. In this talk, I will describe Resilient Composition, a new abstraction that ensures fault-tolerance in applications composed from independent, distributed components. The key insight is to rely on atomic, fault-tolerant “steps” that span component operations and messages. I will present DARQ, an efficient execution engine for such steps, and Distributed Speculative Execution, a transparent optimization that dramatically reduces the overhead of Resilient Composition. Together, these solutions represent an important step toward a more declarative cloud, where strong primitives separate user applications from their underlying infrastructure.

Bio

Tianyu Li is an incoming Assistant Professor at the University of Wisconsin-Madison. His research interests span distributed systems and database systems, with an emphasis on cloud-native applications. His current work focuses on designing the next generation of cloud runtime that will allow for flexibility and scalability without sacrificing fault-tolerance or performance. Previously, Tianyu obtained his PhD from MIT, and his BS/MS from CMU.