Contributing to Numba¶
We welcome people who want to make contributions to Numba, big or small! Even simple documentation improvements are encouraged. If you have questions, don’t hesitate to ask them (see below).
Numba uses Gitter for public real-time chat. To help improve the signal-to-noise ratio, we have two channels:
- numba/numba: General Numba discussion, questions, and debugging help.
- numba/numba-dev: Discussion of PRs, planning, release coordination, etc.
Both channels are public, but we may ask that discussions on numba-dev move to the numba channel. This is simply to ensure that numba-dev is easy for core developers to keep up with.
Note that the Github issue tracker is the best place to report bugs. Bug reports in chat are difficult to track and likely to be lost.
The core Numba developers have a weekly video conference to discuss roadmap, feature planning, and outstanding issues. These meetings are invite only, but minutes will be taken and will be posted to the Numba wiki.
We use the Github issue tracker to track both bug reports and feature requests. If you report an issue, please include specifics:
- what you are trying to do;
- which operating system you have and which version of Numba you are running;
- how Numba is misbehaving, e.g. the full error traceback, or the unexpected results you are getting;
- as far as possible, a code snippet that allows full reproduction of your problem.
Getting set up¶
If you want to contribute, we recommend you fork our Github repository, then create a branch representing your work. When your work is ready, you should submit it as a pull request from the Github interface.
If you want, you can submit a pull request even when you haven’t finished
working. This can be useful to gather feedback, or to stress your changes
against the continuous integration platform. In this
case, please prepend
[WIP] to your pull request’s title.
Numba has a number of dependencies (mostly NumPy and llvmlite) with non-trivial build instructions. Unless you want to build those dependencies yourself, we recommend you use conda to create a dedicated development environment and install precompiled versions of those dependencies there.
First add the Anaconda Cloud
numba channel so as to get development builds
of the llvmlite library:
$ conda config --add channels numba
Then create an environment with the right dependencies:
$ conda create -n numbaenv python=3.6 llvmlite numpy scipy jinja2 cffi
This installs an environment based on Python 3.6, but you can of course
choose another version supported by Numba. To test additional features,
you may also need to install
To activate the environment for the current shell session:
$ conda activate numbaenv
These instructions are for a standard Linux shell. You may need to adapt them for other platforms.
Once the environment is activated, you have a dedicated Python with the required dependencies:
$ python Python 3.6.6 |Anaconda, Inc.| (default, Jun 28 2018, 11:07:29) [GCC 4.2.1 Compatible Clang 4.0.1 (tags/RELEASE_401/final)] on darwin Type "help", "copyright", "credits" or "license" for more information. >>> import llvmlite >>> llvmlite.__version__ '0.24.0'
For a convenient development workflow, we recommend you build Numba inside its source checkout:
$ git clone git://github.com/numba/numba.git $ cd numba $ python setup.py build_ext --inplace
This assumes you have a working C compiler and runtime on your development system. You will have to run this command again whenever you modify C files inside the Numba source tree.
Numba is validated using a test suite comprised of various kind of tests
(unit tests, functional tests). The test suite is written using the
The tests can be executed via
python -m numba.runtests. If you are
running Numba from a source checkout, you can type
as a shortcut. Various options are supported to influence test running
and reporting. Pass
--help to get a glimpse at those options.
to list all available tests:
$ python -m numba.runtests -l
to list tests from a specific (sub-)suite:
$ python -m numba.runtests -l numba.tests.test_usecases
to run those tests:
$ python -m numba.runtests numba.tests.test_usecases
to run all tests in parallel, using multiple sub-processes:
$ python -m numba.runtests -m
For a detailed list of all options:
$ python -m numba.runtests -h
The numba test suite can take a long time to complete. When you want to avoid the long wait, it is useful to focus on the failing tests first with the following test runner options:
--failed-firstoption is added to capture the list of failed tests and to re-execute them first:
$ python -m numba.runtests --failed-first -m -v -b
--last-failedoption is used with
--failed-firstto execute the previously failed tests only:
$ python -m numba.runtests --last-failed -m -v -b
When debugging, it is useful to turn on logging. Numba logs using the
logging module. One can use the standard ways (i.e.
logging.basicConfig) to configure the logging behavior. To enable logging
in the test runner, there is a
--log flag for convenience:
$ python -m numba.runtests --log
Any non-trivial change should go through a code review by one or several of the core developers. The recommended process is to submit a pull request on github.
A code review should try to assess the following criteria:
- general design and correctness
- code structure and maintainability
- coding conventions
- docstrings, comments
- test coverage
All Python code should follow PEP 8. Our C code doesn’t have a well-defined coding style (would it be nice to follow PEP 7?). Code and documentation should generally fit within 80 columns, for maximum readability with all existing tools (such as code review UIs).
Numba uses Flake8 to ensure a consistent
Python code format throughout the project.
flake8 can be installed
conda and then run from the root of the Numba repository:
Optionally, you may wish to setup pre-commit hooks
to automatically run
flake8 when you make a git commit. This can be
done by installing
pip install pre-commit
and then running:
from the root of the Numba repository. Now
flake8 will be run each time
you commit changes. You can skip this check with
git commit --no-verify.
Numba has started the process of using type hints in its code base. This will be a gradual process of extending the number of files that use type hints, as well as going from voluntary to mandatory type hints for new features. Mypy is used for automated static checking.
At the moment, only certain files are checked by mypy. The list can be found in
mypy.ini. When making changes to
those files, it is necessary to add the required type hints such that mypy tests will pass. Only in exceptional
type: ignore comments be used.
If you are contributing a new feature, we encourage you to use type hints, even if the file is not currently in the
checklist. If you want to contribute type hints to enable a new file to be in the checklist, please add the file to the
files variable in
mypy.ini, and decide what level of compliance you are targetting. Level 3 is basic static
checks, while levels 2 and 1 represent stricter checking. The levels are described in details in
There is potential for confusion between the Numba module
typing and Python built-in module
typing used for type
hints, as well as between Numba types—such as
typing types of the same name.
To mitigate the risk of confusion we use a naming convention by which objects of the built-in
typing module are
imported with an
pt prefix. For example,
typing.Dict is imported as
from typing import Dict as ptDict.
master branch is expected to be stable at all times.
This translates into the fact that the test suite passes without errors
on all supported platforms (see below). This also means that a pull request
also needs to pass the test suite before it is merged in.
Every commit to the master branch is automatically tested on all of the platforms Numba supports. This includes ARMv7, ARMv8, POWER8, as well as both AMD and NVIDIA GPUs. The build system however is internal to Anaconda, so we also use Travis CI and Azure to provide public continuous integration information for as many combinations as can be supported by the service. Travis CI automatically tests all pull requests on OS X and Linux, as well as a sampling of different Python and NumPy versions, Azure does the same but also includes Windows. If you see problems on platforms you are unfamiliar with, feel free to ask for help in your pull request. The Numba core developers can help diagnose cross-platform compatibility issues.
The Numba documentation is split over two repositories:
- This documentation is in the
docsdirectory inside the Numba repository.
- The Numba homepage has its sources in a separate repository at https://github.com/numba/numba-webpage
To build the documentation, you need the bootstrap theme:
$ pip install sphinx_bootstrap_theme
You can edit the source files under
docs/source/, after which you can
build and check the documentation:
$ make html $ open _build/html/index.html
Core developers can upload this documentation to the Numba website
at http://numba.pydata.org by using the
gh-pages.py script under
$ python gh-pages.py version # version can be 'dev' or '0.16' etc
then verify the repository under the
gh-pages directory and use
Web site homepage¶
After pushing documentation to a new version, core developers will want to update the website. Some notable files:
index.rst# Update main page
_templates/sidebar_versions.html# Update sidebar links
doc.rst# Update after adding a new version for numba docs
download.rst# Updata after uploading new numba version to pypi
After updating run:
$ make html
and check out
_build/html/index.html. To push updates to the Web site:
$ python _scripts/gh-pages.py
then verify the repository under the
gh-pages directory. Make sure the
CNAME file is present and contains a single line for
git push to update the website.