Zeno Does Data Science- The Paradoxical Quest for Reproducibility
summary
Describes concrete fixes: adopt one virtual environment per repo, use git for readable text artifacts only, separate a “what I want” env YAML from a locked environment (via conda lock workflows), script setup with makefiles and numbered notebooks, hide local paths in a config file and keep secrets out with dotenv, run iterative workshops and track failures. Claims a measurable goal—15 minutes start-up target—and reports hitting it for most attendees after these changes, trading perfect purity for a pragmatic reproducible baseline.
AI-written summary · about
chapters (16)
- 0:00Reproducibility in data science
- 1:27Zeno’s Paradox analogy
- 4:35The failure that sparked journey
- 6:18Cookiecutter data science
- 7:43Wrong goal for workflows
- 9:27Where reproducibility begins
- 11:10New methods threshold
- 13:35Six lessons from workshops
- 15:20Lesson 1: Using git
- 18:47Lesson 2: Virtual environments
- 22:39Lesson 3: Lock files
- 26:28Lesson 4: Scripting workflow
- 29:21Lesson 5: Eliminating state
- 35:42Lesson 6: Breadcrumbs for users
- 37:28Defining reproducibility success
- 38:51Solving Zeno’s Paradox
related talks
23mReproducible Environments for Reproducible Results
30mReproducible work environments for data scientists using Nix
20mBetter Care and Feeding of Machine Learning Models
30mReproducible Machine Learning
33mBuilding Reproducible ML Processes with an Open Source Stack
39mNotebook To Production
15mReproducible research using Bazel
47mWithout Data, you're just another Person with Opinions