Scale Interoperable Composite Backend Services With AFCL Workflows in Serverless Sky Computing

Cloud providers offer serverless, pre-trained ML-based Backend-as-a-Service (BaaS) services, for tasks like translation or speech conversion via single API calls. However, for more complex conversions, users compose multiple BaaS services into serverless workflows to overcome individual limitations. This paper investigates the scalability and interoperability of such composite BaaS services within Sky Computing, focusing on dynamic provider selection between AWS and GCP. We develop two scalable composite BaaS services and propose a novel approach to federate BaaS services, offering two interoperability mechanisms: (1) through-workflow and (2) through-function, enabling seamless integration and dynamic selection of cloud providers. Experiments reveal significant runtime variability in Sky Computing, with extreme outliers of up to $12.9 \times $ , making naive averaging strategies unsuitable to determine the optimal configuration. In addition to the well-known cold and warm start effects in FaaS performance models, we identify four additional categories of anomalies that, when FaaS and BaaS are combined, generate extreme outliers far beyond baseline performance instability. Finally, we propose a theoretical framework to explain these behaviors and mitigate their impact, paving the way for reliable and scalable composite BaaS workflows.