Description
I was thinking particularly of cases like datetime64, Categorical and GeoSeries (from geopandas).
I recently posted on the numpy discussion mailing to try to get a sense of what solutions exist for writing custom dtypes without writing C. Unfortunately, it appears there's not much hope!
http://mail.scipy.org/pipermail/numpy-discussion/2014-September/071231.html
@njsmith suggested that I really should ask pandas developers to chime in to find out why they choose to work around numpy's limitations rather than enhance it. I would love it if someone who understands the "why" for the choices pandas made could add their perspective to that thread.
@jreback @JanSchulz any thoughts to add? or did a sum it up well enough with "nobody wants to write C"?