AexPy: Detecting API Breaking Changes in Python Packages

With the popularity of the Python language, com-munity developers create and maintain a lot of third-party packages. APIs change frequently during the package evolving. Package developers need keeping backward compatibility of APIs to avoid breaking client code. Detecting breaking changes in Python packages is challenging because of Python's dynamic features and flexible designs, such as dynamic typing, API aliases, confusing public boundary and flexible argument passing. Despite the language's popularity, there have been few tools aiming to detect breaking changes, and existing approaches lack sufficient consideration of the mentioned challenges, resulting in imprecise and incomplete detection. Briefly, we propose an API-model-based systematic approach to address this problem. We design a Python-specific API model and classify API changes in different breaking levels. Based on the model, we obtain APIs with their types by a robust hybrid analysis and detect graded changes by checking constraints on API pairs. We implement a prototype, AexPy. Thanks to the more comprehensive model and hybrid analysis adopted, AexPy outperforms the state-of-the-art techniques for Python with an 86.9% recall on a dataset of 61 known breaking changes. Besides, AexPy detects 405 (manually verified) high and medium breaking changes with a precision of 93.5% on the latest versions of 45 packages. Specifically, in addition to 291 documented breaking changes, AexPy detects 114 undocumented changes. We report 63 undocumented breaking changes to active package developers, and 31 have been confirmed.