ML-FaaS: Toward Exploiting the Serverless Paradigm to Facilitate Machine Learning Functions as a Service

Abstract

Serverless computing has emerged as a revolutionary model that enables the deployment of applications and services by raising the level of abstraction from the underline resources. Its main functionality is enlightened by the notion of Function-asa-Service (FaaS) as the core means to realize efficient serverless offerings. Following the shift from traditional architectures to microservices -by attaining flexibility, productivity, portability, and performance in industrial-scale IT projects, the serverless model introduces even more fine-grained services, named “nanoservices”, which facilitate required scalability by abstracting the deployment and management of the infrastructure resources. On the application space, advances in big data analysis contribute towards extracting actionable knowledge in various application domains. In this context, approaches for big data analysis aim at exploiting the added value of serverless architectures. To this end, we are presenting an extendable and generalized approach for facilitating the provision of Machine Learning Functions-as-a-Service (MLFaaS). The proposed approach outstrips the classical atomic and standard isolated services by facilitating composite services, i.e., workflows/pipelines of ML tasks, thus enabling the realization of the complete data path functions as required by data scientists. We demonstrate the operation of the proposed approach by modeling a real-world analytics scenario as an ML workflow pipeline and evaluate its performance in terms of performance. Furthermore, we address the challenge of utilizing a function oriented service template recommendation system, by expanding the serverless functional boundaries towards a holistic Quality-of-Service (QoS)-aware service function selection approach based on Artificial Intelligence techniques. These techniques propose the optimal number of functions to be implemented in a pipeline by exploiting the importance of response time as the primary key of the application’s performance.