Maintainability Definition, Scoping, and Measurement for Machine Learning Systems

Other conference contribution, 2025

Machine learning (ML) systems are increasingly being applied to critical tasks. Like all systems, ML systems must satisfy maintainability requirements. However, assessing and ensuring maintainability is complicated by the dynamic, heterogeneous, and interconnected components within ML systems—a mixture of structured code, scripting, data, and models. We propose that ensuring maintainability of ML systems requires definition, specification, and measurements that can be scoped across one or more of these components, in addition to the system as a whole. To that end, we propose a component-based breakdown of ML systems, a modified definition of maintainability, and examples of modified modularity measurements. We use these to characterize the modularity of real-world ML systems. Our contributions offer a starting point for future research on maintainability for ML systems.