Restoring Execution Environments of Jupyter Notebooks

Reviewed by Greg Wilson / 2023-03-23
Keywords: Computational Notebooks, Scientific Computing

Like the paper reviewed yesterday, this one looks at computational notebooks: more specifically, at how to reverse engineer their execution environment. As the abstract says, more than 90% of published Jupyter notebooks don't explicitly state package dependencies, which can make them non-reproducible at best and non-executable in far too many cases. Wang et al. developed a tool that collects package APIs, analyzes notebooks to determine which ones are needed, and then finds combinations of packages that will make the notebook run. Quoting from the paper, "In a lab setting, SnifferDog is effective in automatically inferring execution environments for Jupyter notebooks, successfully generating installation requirements for 315/340 (92.6%) of notebooks. 284/315 (90.2%) of notebooks could be executed automatically."

This is impressive work, and the authors have made it available on GitHub. Once again, I hope it will inform the design of a new generation of notebooks: rather than requiring people to try to scrape dependencies, for example, I hope that future notebook systems will capture them automatically.

Jiawei Wang, Li Li, and Andreas Zeller. Restoring execution environments of Jupyter notebooks. 2021. arXiv:2103.02959.

More than ninety percent of published Jupyter notebooks do not state dependencies on external packages. This makes them non-executable and thus hinders reproducibility of scientific results. We present SnifferDog, an approach that 1) collects the APIs of Python packages and versions, creating a database of APIs; 2) analyzes notebooks to determine candidates for required packages and versions; and 3) checks which packages are required to make the notebook executable (and ideally, reproduce its stored results). In its evaluation, we show that SnifferDog precisely restores execution environments for the largest majority of notebooks, making them immediately executable for end users.