Multiverse Notebook: Shifting Data Scientists to Time Travelers
summary
Multiverse Notebook implements fork per cell kernel checkpoints so each cell runs in a child process and time-traveling is just forking a new child, avoiding serialization and reverse computation. It adds a CPython-aware optimization that immortalize parent objects by assigning cells to memory regions to reduce copy-on-write blowup. Evaluation replays 10 Kaggle notebooks (median replay 200s, max ~2,000s; median RAM 8 GB, max 60 GB), finds fork checkpointing several times faster than serialization and a median memory reduction of 6.5% (best >50%). It doesn't rewind external files or file descriptors.
AI-written summary · about
chapters (9)
- 0:00Multiverse Notebook intro
- 1:50Code versus kernel state
- 3:58Motivation for time-traveling
- 5:41Research contributions
- 7:09Time travel with fork
- 9:20Python memory challenges
- 12:51Evaluation results
- 14:59Summary and resources
- 15:41Q&A on stateful objects
related talks
33mWorld's First Undoable Notebook
27mThe Future of Notebooks in a Claude Code World
36mRewind: Python Time-Travel Debugger
25mTime Travel Debuggers
28mTemporal Breakpoints for Multiverse Debugging
29mOh no! I think my project is outgrowing my Jupyter notebook. How do I survive?
01h 23mImproving Your Data Modeling Work Through Open-Source Software
25mTo Notebook or Not to Notebook: Multilingual Workflows for Data Analysis