The performance hell thing is also also kind of a virtue, though. The language is awful, so everything that does any amount of compute is FFI'd into third party libraries (numpy, torch, sympy, etc). Those libraries are for the most part pretty well designed... or, at least, keep you in a few pretty well-constrained patterns that are easy enough to translate.

If you've ever read through FORTRAN code from a mathematics department or MATLAB/C/C++ from (non-software) engineering disciplines, then you probably understand why productionizing a jupyter notebook is definitely not the worst of all possible worlds.