Automated Versioning for Software Releases: A Retrospective Study and A New Lightweight Approach

抄録

Third-party libraries are increasingly used in software development. With new versions released, many conflicts due to API breaking may be introduced. Therefore, semantic versioning with a set of rules and requirements is proposed to inform users about incompatibilities and other changes in a new release. Simply speaking, major, minor, and patch version numbers should be updated when breaking, non-breaking, and internal changes are made, respectively. However, many third-party libraries do not follow semantic versioning principles, bringing many efforts to adapt new versions of libraries. Hence developers are often unwilling to update dependencies. In this paper, we first retrospectively investigate the three types of existing techniques that can be used in automatic versioning, a task for more reliable versions: 1) rule-based, considering keywords only, 2) machine learning-based, concerning many aspects of features, and 3) source code analysis for semantic versioning compliance. Having found the limitation of these approaches, we propose a new and simple approach AutoVer, which can capture the intent of developers through commit messages. Specifically, we treat the tokens and phrases in the commit messages as features and train AutoVer using XGBoost, a well-known machine learning method that performs well in many classification tasks. The evaluation results show that AutoVer outperforms our investigated approaches in terms of many metrics. Specifically, the major, minor and major type F1-Scores of AutoVer are 0.889, 0.992 and 0.998, respectively. We also comprehensively investigate many settings that affect the performance of AutoVer, including choice of machine learning model, number of features and keywords in commit messages as well as provide practitioners and researchers with some implications for future studies, e.g., writing clear commit messages for better understanding the intent of making changes, and combining semver-compliance checking and automatic versioning.