David Liu will present his Pre-FPO "A Serverless Architecture for Application-level Orchestration" on Tuesday, May 17, 2022 at 2:00 PM via Zoom.

Zoom link: https://princeton.zoom.us/j/92893120748

Committee Members: Amit Levy (Examiner, advisor), Mike Freedman (Examiner, advisor), Jennifer Rexford(examiner), Wyatt Lloyd (reader), Ravi Netravali (reader).

All are welcome to attend.

Title: A Serverless Architecture for Application-level Orchestration

Abstract:

Serverless workflow orchestrators make it easier for developers to build complex serverless applications by driving the execution of a workflow from a central service.  Making orchestrators fault-tolerant and scalable requires engineering effort as well as large up-front resource costs to deploy. As a result, developers typically rely on multi-tenant services run by cloud providers that can amortize these costs across many users.

Unfortunately, this centralization comes at the expense of application flexibility. Orchestrators support a fixed set of application patterns, and trade off performance, scalability, fault tolerance, and resource efficiency using a policy chosen by the cloud provider. They support many applications well, but cannot simultaneously satisfy the needs of all applications.

We argue that centralized serverless orchestrators are actually unnecessary. Instead, a decentralized orchestrator, deployed as a library included by the serverless application, can support the same application patterns and fault tolerance as existing orchestrators using only the basic building blocks of serverless platforms---a scalable ``Function'' executor (FaaS) and a consistent scalable datastore. This enables different applications to use or customize different libraries to implement different patterns or make application-specific trade-offs.

To demonstrate this, we describe the design and implementation of Unum, a decentralized serverless workflow system. Unum compiles workflows written in existing high-level workflow languages into a partitioned orchestration library that executes orchestration logic in-situ with user functions. Unum provides the same strong correctness guarantees under faults as a centralized orchestrator, performs similarly, scales better, and costs less than existing centralized workflow systems.

Our implementation of Unum supports a superset of applications the AWS Step Functions orchestrator supports, runs on both AWS Lambda and Google Cloud Functions, and runs up to 2x faster and up to 9x cheaper than native Step Functions on a representative set of applications.