Search

Software reproducibility is important for re-usability and thecumulative progress of research. An important manifestation ofunreproducible software is the outcome of software builds changingover time. While enhancing code reuse, the usage of open-sourcedependency packages hosted on centralized software repositories likePyPI can have adverse effects on build reproducibility. Frequentupdates of these packages often cause their latest versions to havebreaking changes for applications using them. Large Pythonapplications risk their historical builds to become unreproducible dueto the widespread usage of Python dependencies, and the lack ofuniform practices for dependency version specification. Manuallyfixing dependency errors requires expensive developer time and effort,while automated approaches face challenges such as parsingunstructured build logs, finding transitive dependencies, and dealingwith an exponential search space of dependency versions. In thisthesis, we investigate how open-source Python projects specifydependency versions, and how their reproducibility is impacted bydependency packages. We propose a tool PyDFix to detect and fixunreproducibility in Python builds caused by dependency versionerrors. The ability of PyDFix to fix unreproducible builds is evaluatedon two bug datasets BugSwarm and BugsInPy, both of which are builtfrom real-world open-source projects. PyDFix analyzes a total of 2,702builds, identifying 1,921 (71.1%) of them to be unreproducible due todependency errors. From these, PyDFix provides a complete fix for 859(44.7%) builds, and partial fixes for an additional 632 (32.9%)builds.