A Comprehensive Multi-Vocal Empirical Study of ML Cloud Service Misuses
Hadil Ben Amor, Manel Abdellatif, Taher A. Ghaleb · ACM Transactions on Software Engineering and Methodology · 2026
Machine Learning (ML) models are widely adopted across many application domains. To support this trend, major cloud providers offer ML cloud services that eliminate the need to build models from scratch. These ML services empower practitioners with varying expertise to rapidly embed ML capabilities into software systems. However, practitioners often overlook best practices and optimal design when using cloud-based services. This results in recurring misuses that can compromise system quality and may hinder long-term maintenance and evolution. Though prior work has studied some misuse cases, the field lacks consistent terminology and clear specifications of ML service misuses. In this paper, we address three research questions (RQs) investigating (1) the types of ML service misuses identified in a systematic review of the research literature and gray literature as well as open-source software systems, (2) the state of the practice of ML cloud service misuses in industry, and (3) the prevalence of ML service misuses in open-source projects in comparison with what is observed in industry practice. To address these RQs, we conducted a comprehensive multi-vocal empirical study investigating the prevalence of ML cloud service misuses in practical contexts. Overall, our analyses incorporated a diverse range of sources, including academic research, official documentation from major ML cloud providers, and 377 GitHub projects using ML services, and were complemented by a survey of 50 ML practitioners from industry. As a result, we developed a catalog of 20 ML service misuses, cross-validated through evidence from both open-source projects and industry feedback. We observe that our identified misuses are widely prevalent in both open-source projects and industry, often due to a lack of understanding of service capabilities, inadequate documentation, and poor awareness of best practices. This highlights the importance of ongoing education on best practices for ML services and highlights the need for tools to automatically detect and refactor ML service misuses.