Multiverse Notebook: Shifting Data Scientists to Time Travelers

Abstract

Computational notebook environments are popular and de facto standard tools for programming in data science, whereas computational notebooks are notorious in software engineering. The criticism there stems from the characteristic of facilitating unrestricted dynamic patching of running programs, which makes exploratory coding quick but the resultant code messy and inconsistent. In this work, we first reveal that dynamic patching is a natural demand rather than a mere bad practice in data science programming on Kaggle. We then develop Multiverse Notebook, a computational notebook engine for time-traveling exploration. It enables users to time-travel to any past state and restart with new code from there under state isolation. We present an approach to efficiently implementing time-traveling exploration. We empirically evaluate Multiverse Notebook on ten real-world tasks from Kaggle. Our experiments show that time-traveling exploration on Multiverse Notebook is reasonably efficient.