Notebooks and reproducibility

Jupyter notebooks mix code, output, and prose—great for EDA and communication. Reproducibility means another person (or future you) can rerun and get the same conclusions.

Notebook strengths and risks

• Strengths — iterative plots, teaching, stakeholder walkthroughs
• Risks — out-of-order execution, hidden state, huge diffs in Git

Reproducibility checklist

1. Pin package versions (requirements.txt or conda env)
2. Set random seeds for splits and models
3. Record data snapshot path or query hash
4. Restart kernel and Run All before sharing
5. Move stable logic to .py modules tested in CI

Git with notebooks

Use nbstripout or review tools; prefer scripts for production pipelines. Notebooks are artifacts; tested functions are products.

Playground vs local

# Local workflow:
# python -m venv .venv && source .venv/bin/activate
# pip install jupyter pandas numpy matplotlib
# jupyter lab

This site’s lessons use server_script; notebooks run on your machine with full PyPI stack.

Important interview questions and answers

1. Q: Why Run All?
   A: Ensures cell order matches saved state—catches variables defined only in later cells.
2. Q: Notebook vs module?
   A: Modules import cleanly in pipelines; notebooks excel for exploration and reports.

Self-check

1. Name three reproducibility practices.
2. What risk comes from out-of-order notebook execution?
3. Why extract logic to .py files for production?

Tip: Pin versions in requirements.txt or environment.yml.