Sarcouncil Journal of Engineering and Computer Sciences

Abstract: This article presents a novel predictive and adaptive scheduling framework designed to optimize serverless-based machine learning pipelines for both cost efficiency and low-latency execution. The article addresses critical challenges in serverless ML deployments through three key innovations: intelligent cold start prediction models that anticipate function initialization delays, dynamic directed acyclic graph (DAG) reshaping strategies that adapt workflow structure in response to changing conditions, and a transient state orchestration layer that efficiently manages data transfer between ephemeral functions. The article demonstrates that this integrated approach significantly reduces end-to-end latency and operational costs while maintaining high reliability and throughput. The article automatically navigates complex trade-offs between competing objectives, adapting to workload characteristics and user priorities without requiring manual optimization. By enabling more efficient execution of ML pipelines in serverless environments, this article challenges prevailing assumptions about serverless computing's limitations for complex workflows. It offers ML practitioners a powerful new paradigm for deploying scalable, cost-effective machine learning systems in the cloud. The article generalizes across major serverless platforms and demonstrates particular strength in handling bursty workloads and complex computational graphs typical of modern machine learning applications.