When code smells meet ML: on the lifecycle of ML-specific code smells in ML-enabled systems

The adoption of Machine Learning (ML)-enabled systems is growing rapidly, introducing novel challenges in maintaining quality and managing technical debt in these complex systems. Among the key quality threats are ML-specific code smells (ML-CSs), suboptimal implementation practices in ML pipelines that can compromise system performance, reliability, and maintainability. Although these smells have been defined in the literature, detailed insights into their characteristics, evolution, and mitigation strategies are still needed to help developers address these quality issues effectively. In this paper, we investigate the emergence and evolution of ML-CSs through a large-scale empirical study focusing on (i) their prevalence in real ML-enabled systems, (ii) how they are introduced and removed, and (iii) their survivability. We analyze over 400,000 commits from 337 ML-enabled projects, leveraging CodeSmile, a novel ML smell detector that we developed to enable our investigation and identify ML-specific code smells. Our results reveal that: (1) CodeSmile can detect ML-CSs with precision and recall rates of 87.4% and 78.6%, respectively; (2) ML-CSs are frequently introduced during file modifications in new feature tasks; (3) smells are typically removed during tasks related to new features, enhancements, or refactoring; and (4) the majority of ML-CSs are resolved within the first 10% of commits. Based on these findings, we provide actionable conclusions and insights to guide future research and quality assurance practices for ML-enabled systems.