Releasing to PyPi
Category : Python
While attempting to explain the Datascroller release process to collaborators John and Kevin over Hangouts, it quickly became apparent that these steps needed to be put in writing. The process below is very manual, tedious, and constructed from a trial and error approach. So take with a grain of salt and leave a comment if you know a better way!
Verifying the code to be released
PyPi is not especially forgiving if you release code you didn’t intend to, so it’s worth a bit of time to double check. After checking out the master branch and pulling changes from the remote, I check the following:
versionin setup.py the right one for the release?
- If using
download_url(which I just learned is not advised), does the suffix align with the version?
After confirming that the right code is in my local repo, the next checks concern the functionality. Datascroller only has one test script right now that is not hooked up to CI/CD, so here are our steps:
- Create a virtual environment with venv (e.g.,
python -m venv myenvs/test) and enter it
- While still in my local repo, perform a developer install using pip:
pip install -e .
- Run any test scripts you have
For Datascroller I like to repeat the steps above on both Windows and Linux.
Uploading to TestPyPi
With PyPi, if you successfully release version 1.2.0 and then realize something is wrong with the code, then that’s too bad because you’ll never be able to release it again. You can release version 1.2.1, but that requires a code change to setup.py and might mess up your plans. That’s where TestPyPi comes in (make an account if you don’t have one). Back in the root of your local repo, while still in the virtual environment created above, perform the following steps:
pip install twine(a package especially for publishing to PyPi)
python setup.py sdist(creates the “dist” source distribution folder)
twine upload --skip-existing --repository-url https://test.pypi.org/legacy/ dist/*
You’ll then get a link to go look at your package on TestPyPi. “See, looks good!” I told collaborators John and Kevin. “So, what exactly looks good?” asked John. He had a point, so I created new virtual environments and ran
to ensure a key demo worked when installed through TestPyPi. Note in the command above that you have to link to the real PyPi to get third-party packages such as Pandas. This is enabled via the flag
If something when wrong on version 1.2.0, then even on TestPyPi, you cannot delete and resubmit version 1.2.0. You would need to change the version in setup.py and then change it back before submitting to PyPi. In these cases I would append a suffix to the current version (e.g., 1.2.0rc1).
Uploading to PyPi
If everything looks good and your account is in good standing on PyPi.org, then you’re ready for the very last step:
twine upload --skip-existing dist/*
Getting right with Github
After a successful release to PyPi, the final step is to go to the Releases section of the Datascroller project and draft a new release. This will snapshot the project and provide downloads of the entire codebase in zip and tar.gz form. Since we were using the
download_url argument in setup.py, we’d try to match the future name of the tarball that only exists after this step. Provided we really can (or should) leave that argument out next time, that’s one less thing to worry about.
When showing this workflow to John and Kevin, the feedback was, “there has to be a way to automate this.” Creating multiple virtual environments and installations in itself a hassle, so this is something we’ll be investing in soon. Please comment below regarding any weak points in the process above and suggestions to make releasing to PyPi easier and more fun!