Infrastructure Guide¶
This guide serves as an advanced version of the contributing documentation and contains the information on how we manage MetPy and the infrastructure behind it.
Versioning¶
To manage identifying the version of the code, MetPy relies upon setuptools_scm. setuptools_scm
takes the current version of the
source from git tags and any additional commits. For development, the version will have a
string like 0.10.0.post209+gff2e549f.d20190918
, which comes from git describe
. This
version means that the current code is 209 commits past the 0.10.0 tag, on git hash
ff2e549f
, with local changes on top, last made on a date (indicated by d20190918
). For
a release, or non-git repo source dir, the version will just come from the most recent tag
(i.e. v0.10.0
).
To make a new version, simply add a new tag with a name like vMajor.Minor.Bugfix
and push
to GitHub. Github will add a new release with a source archive.zip file. Running
python setup.py sdist
will build a new source distribution with the appropriately generated version file as well. This will also create a new stable set of documentation.
Documentation¶
MetPy’s documentation is built using sphinx >= 1.8. API documentation is automatically
generated from docstrings, written using the
NumPy docstring standard.
There are also example scripts in the examples
directory. Using the sphinx-gallery
extension, these examples are executed and turned into a gallery of thumbnails. The
extension also makes these scripts available as Jupyter notebooks.
The documentation is hosted on GitHub Pages. The docs are
built automatically from master
with every build on Travis-CI; every merged PR will
have the built docs upload to GitHub Pages. As part of the build, the documentation is also
checked with doc8
. To see what the docs will look like, you also need to install the
sphinx-rtd-theme
package.
Other Tools¶
Continuous integration is performed by Travis CI
and AppVeyor.
Travis runs the unit tests on Linux for all supported versions of Python, as well as runs
against the minimum package versions. flake8
(with the pep8-naming
and
flake8-quotes
plugins) is also run against the code to check formatting. Travis is also
used to build the documentation and to run the examples to ensure they stay working. AppVeyor
is a similar service; here the tests and examples are run against Python 2 and 3 for both
32- and 64-bit versions of Windows.
Test coverage is monitored by codecov.io.
The following services are used to track code quality:
Releasing¶
To create a new release, go to the GitHub page and make a new release. The tag should be a sensible version number, like v1.0.0. Add a name (can just be the version) and add some release notes on what the big changes are. It’s also possible to use loghub to get information on all the issues and PRs that were closed for the relevant milestone.
PyPI¶
Once the new release is published on GitHub, this will create the tag, which will trigger new builds on Travis (and AppVeyor, but that’s not relevant). When the main test build on Travis (currently Python 3 tests) succeeds, Travis will handle building the source distribution and wheels, and upload them to PyPI.
To build and upload manually (if for some reason it is necessary):
Do a pull locally to grab the new tag. This will ensure that
setuptools_scm
will give you the proper version.(optional) Perform a
git clean -f -x -d
from the root of the repository. This will delete everything not tracked by git, but will also ensure clean source distribution.MANIFEST.in
is set to include/exclude mostly correctly, but could miss some things.Run
python setup.py sdist bdist_wheel
(this requires thatwheel
is installed).Upload using
twine
:twine upload dist/*
, assuming thedist/
directory contains only files for this release. This upload process will include any changes to theREADME
as well as any updated flags fromsetup.py
.
Conda¶
MetPy conda packages are automatically produced and uploaded to Anaconda.org thanks to conda-forge. Once the release is built and uploaded to PyPI, then a Pull Request should be made against the MetPy feedstock, which contains the recipe for building MetPy’s conda packages. The Pull Request should:
Update the version
Update the hash to match that of the new source distribution uploaded to PyPI
Reset the build number to 0 (if necessary)
Update the dependencies (and their versions) as necessary
The Pull Request will test building the packages on all the platforms. Once this succeeds, the Pull Request can be merged, which will trigger the final build and upload of the packages to anaconda.org.