[Numpy-discussion] Latest Array-Interface PEP

Perry Greenfield perry at stsci.edu
Tue Jan 9 13:37:56 EST 2007

A few comments regarding what I think justifies some sort of standard  
to be part of Python (understanding that there are various ways that  
it could be done, so I'm not commenting on the specifics here directly).

I don't there is any harm in making the standard numpy-centric. In  
fact, I think the selling point is that this standard means that any  
extension can expose repetitive data to those that want to manipulate  
it *in Python* in a simple and standard way. While its possible that  
ultimately there are those that will pass these sorts of things from  
one extension into another, I don't see that as a common use for a  
while. What it does mean is that if you want a simple Python-only way  
of seeing and modifying such data, all you need to do is install  
numpy. You don't have write a C extension. If the extension writers  
use the descriptive field names, they can make the array they expose  
somewhat self-documenting so that turning it into an array and doing  
a little introspection may tell you all you need to know about the  
data. Rather than try to sell this as some neutral interface, I would  
make the numpy-dependence explicit (not that it excludes extension-to- 
extension direct use).

It may be that the developers of various extensions are not the ones  
most interested in this capability (after all, they've build it to do  
what they want), but I wouldn't be surprised if many of the users of  
the extension would like it so they can do things the extension  
doesn't allow them. So one approach is to see what the respective  
user communities think about such capabilities. If they find out what  
this can do for them, they may pressure the developers for such  
support. People in the numpy community can also volunteer to  
implement the standard (but with this approach, it's a bit of a  
chicken and egg thing as someone has mentioned. You can't do it if it  
isn't in Python yet.)

I do agree that the most persuasive approach would be to have at  
least some of the 3rd party extensions support this explicitly on the  
python-dev list.


More information about the NumPy-Discussion mailing list