Leveraging External Synchrony for Distributed Databases in Functions as a Service Workflows

Abstract

Functions-as-a-Service (FaaS) is a rapidly growing cloud computing paradigm. It promises efficient utilization of resources and removes the need for developers to manage their own infrastructure. FaaS workflows must use globally distributed databases to store data persistently across the workflow. Furthermore, to guarantee the consistency of data in the workflow, FaaS workflows must block for globally distributed database writes to become durable. This substantially increases the end-to-end latency for FaaS workflows. This paper takes a novel approach to addressing this issue by leveraging the concept of external synchrony. The external synchrony protocol is causally consistent and guarantees that the workflow computation will be indistinguishable from the baseline synchronous procedure to an external observer of the FaaS workflow, yet the external synchrony protocol also significantly improves performance. This paper finds that the external synchrony protocol empirically improves performance in synthetic single function microbenchmarks and also provides a 21.03% improvement in a real FaaS workflow. These results demonstrate that external synchrony has the potential to significantly reduce FaaS latency and thus provide tangible benefits for developers and users of FaaS.