I find the opposite to be true. With numba & Cython, it’s ridiculously simple to target a segment of Python code for compiled specialization, often required for custom pieces of work not directly expressible in NumPy, especially when processing non-ndarray data structures.

I find I am frequently removing sections of inscrutable “vectorized” NumPy code that is inefficient, not because NumPy is inefficient but because conforming the data to a structure amenable to NumPy is too inefficient in the first place, and rewriting in straightforward, simple Cython functions.