All you need to know is the title is correct. Skip this needlessly verbose and awkwardly argued article. It describes simple variable assignment in REPLs as "caching" (which by conventional software notions, it's not). Anyone who's used a REPL knows to be careful with what's been executed thus far, and the order of variable assignments.

Hopefully any statistician or pydev with 1-2 years of experience figures out a json-based, git-incompatible obstructionist ecosystem does not good science make. Jupyter (nee ipython) has only ever been good for teaching and kaggle-style EDA. It's insane to carry EDA code directly into production (which is why misguided efforts like papermill are doomed to fail).