NEP 36 — Fair play#
Stéfan van der Walt <email@example.com>
This document sets out Rules of Play for companies and outside developers that engage with the NumPy project. It covers:
Restrictions on use of the NumPy name
How and whether to publish a modified distribution
How to make us aware of patched versions
Companies and developers will know after reading this NEP what kinds of behavior the community would like to see, and which we consider troublesome, bothersome, and unacceptable.
Every so often, we learn of NumPy versions modified and circulated by outsiders. These patched versions can cause problems for the NumPy community (see, e.g.,  and ). When issues like these arise, our developers waste time identifying the problematic release, locating alterations, and determining an appropriate course of action.
In addition, packages on the Python Packaging Index are sometimes named such that users assume they are sanctioned or maintained by NumPy. We wish to reduce the number of such incidents.
During a community call on October 16th, 2019 the community resolved to draft guidelines to address these matters.
This document aims to define a minimal set of rules that, when followed, will be considered good-faith efforts in line with the expectations of the NumPy developers.
Our hope is that developers who feel they need to modify NumPy will first consider contributing to the project, or use one of several existing mechanisms for extending our APIs and for operating on externally defined array objects.
When in doubt, please talk to us first. We may suggest an alternative; at minimum, we’ll be prepared.
Fair play rules#
Do not reuse the NumPy name for projects not developed by the NumPy community.
At time of writing, there are only a handful of
numpy-named packages developed by the community, including
unumpy. We ask that external packages not include the phrase
numpy, i.e., avoid names such as
To be clear, this rule only applies to modules (package names); it is perfectly acceptable to have a submodule of your own library named
NumPy is a trademark owned by NumFOCUS.
Do not republish modified versions of NumPy.
Modified versions of NumPy make it very difficult for the developers to address bug reports, since we typically do not know which parts of NumPy have been modified.
If you have to break this rule (and we implore you not to!), then make it clear in the
__version__tag that you have modified NumPy, e.g.:
>>> print(np.__version__) '1.17.2+mycompany.15`
We understand that minor patches are often required to make a library work inside of a distribution. E.g., Debian may patch NumPy so that it searches for optimized BLAS libraries in the correct locations. This is acceptable, but we ask that no substantive changes are made.
Do not extend or modify NumPy’s API.
If you absolutely have to break rule two, please do not add additional functions to the namespace, or modify the API of existing functions. NumPy’s API is already quite large, and we are working hard to reduce it where feasible. Having additional functions exposed in distributed versions is confusing for users and developers alike.
DO use official mechanism to engage with the API.
Protocols such as __array_ufunc__ and __array_function__ were designed to help external packages interact more easily with NumPy. E.g., the latter allows objects from foreign libraries to pass through NumPy. We actively encourage using any of these “officially sanctioned” mechanisms for overriding or interacting with NumPy.
If these mechanisms are deemed insufficient, please start a discussion on the mailing list before monkeypatching NumPy.
Questions and answers#
Q: We would like to distribute an optimized version of NumPy that utilizes special instructions for our company’s CPU. You recommend against that, so what are we to do?
A: Please consider including the patches required in the official NumPy repository. Not only do we encourage such contributions, but we already have optimized loops for some platforms available.
Q: We would like to ship a much faster version of FFT than NumPy provides, but NumPy has no mechanism for overriding its FFT routines. How do we proceed?
A: There are two solutions that we approve of: let the users install your optimizations using a piece of code, such as:
from my_company_accel import patch_numpy_fft patch_numpy_fft()
or have your distribution automatically perform the above, but print a message to the terminal clearly stating what is happening:
We are now patching NumPy for optimal performance under MyComp Special Platform. Please direct all bug reports to https://mycomp.com/numpy-bugs
If you require additional mechanisms for overriding code, please discuss this with the development team on the mailing list.
Q: We would like to distribute NumPy with faster linear algebra routines. Are we allowed to do this?
A: Yes, this is explicitly supported by linking to a different version of BLAS.
References and footnotes#
This document has been placed in the public domain.