Why do Machine Learning Notebooks Crash?

Jupyter notebooks have become central in data science, integrating code, text and output in a flexible environment. With the rise of machine learning (ML), notebooks are increasingly used for ML prototyping and data analysis. However, due to their dependence on complex ML libraries and the flexible notebook semantics that allow cells to be run in any order, notebooks are susceptible to software bugs that may lead to program crashes. This paper presents a comprehensive empirical study focused on crashes in ML notebooks. We collect 64,031 ML notebooks containing 92,542 crashes from GitHub and Kaggle, and manually analyze a sample of 746 crashes across various aspects, including exception types and root causes. Our analysis highlights unique root causes related to notebook semantics, including out-of-order execution and previous cell error, that have not been thoroughly covered in earlier research. Furthermore, we categorize crashes as ML bugs or general Python bugs and examine how the crashes are distributed across different stages of the ML pipeline. Our analysis reveals that 87% of crashes are caused by API misuse, data confusion, notebook-specific issues, environment problems, and implementation errors. Crashes are more commonly related to ML bugs than general Python bugs, particularly in Kaggle notebooks, where over 70% of crashes are ML-related. The distribution of crashes across ML pipeline stages differs between the two platforms. Additionally, most crashes (58%) occur during data preparation, model training, and evaluation/prediction stages of the ML pipeline. GitHub and Kaggle exhibit different crash distributions across these stages.
Comment: submitted to a journal