This a video from a recent conference on Python in Datascience:
It’s still my opinion that using numpy, scipy and pandas in Blender is very useful, and should be done more. “Numeric”, multidimensional arrays are a very good abstraction of most of the data that Blender is dealing with, especially because today’s meshes and particle systems tend to get bigger and bigger. The “dissect a linked data structure” paradigm which dominates the Blender Python API can sometimes get in the way.
Numba is one of many exciting developments in this area. Cython and Theano are two other very nice approaches to optimizing Python code.