Automating Python package release process

I automated this for one of my open-source tools. It‘s all in the file manage.py: https://github.com/tfeldmann/organize/blob/master/manage.py

My project uses poetry for uploading to pypi, so this looks a bit different but should be a good starting point. It also manages the changelog and creates all releases.

python manage.py version steps:

• prompts for version number and checks validity
• updates the __version__.py file
• updates the pyproject.toml file used by poetry
• searches for a section ## WIP in changelog.md and replaces it with current version and todays date.

python manage.py publish steps:

• reads the current version
• reads the changes listed for this version from the changelog
• creates a git tag
• pushes to github (with tags)
• builds and publishes to pypi
• creates a github release with the version number as name and the changes from the changelog as description

The scripts asks for confirmation for each step so things don't get out of hand and prompts for your github and pypi passwords as soon as they are needed.

The following is my own opinionated take on your list. There is a certain range of automation you can achieve, and I'll try to provide a reasonable starting point, and then some hints on how you can go further from there.

---

CI without CD

Adopting this part should already get rid of most of the annoying manual work, and you can automate away more and more as the need arises. If you're not comfortable maintaining a good amount of CI code, you should start here.

Things you'll need are a CI (as you already noted) and a package manager. Something you won't get around is pushing your changes and a new tag with git, so parts of step 5 and 6 remain manual.

Package management

I'll use poetry to keep things concise and because I like it^[1], but there are also other options. This will take care of steps 2, 3, 7, 8, and the unlisted step 10, "update my dependencies and test them for compatibility", which is incredibly annoying as soon as it turns out to be a problem.

The bad news when using poetry is that you'll need to move all packaging configuration into a new file, pyproject.toml. The good news is, that you don't need a separate setup.py, setup.cfg, MANIFEST.in, or requirements.txt any more, since pyproject.toml is a provisional standard for packaging and other tools, and poetry also has a walkthrough on how to port over all the relevant info.

Once the setup is ready, the new deployment workflow would be:

$ poetry update           # update dependencies, may be skipped 
$ poetry version          # bump version
Bumping version from 1.1.2 to 1.1.3
# finalize git stuff, e.g. add -u, commit -m 'v1.1.3', tag v1.1.3, push
$ poetry publish --build  # build and publish to PyPI
Building my_django_lib (1.1.3)
 - Building sdist
 - Built my_django_lib-1.1.3.tar.gz

 - Building wheel
 - Built my_django_lib-1.1.3-py3-none-any.whl

Publishing my_django_lib (1.1.3) to PyPI
 - Uploading my_django_lib-1.1.3-py3-none-any.whl 100%
 - Uploading my_django_lib-1.1.3.tar.gz 100%

This should already be a lot shorter than what you're currently doing. If you always execute the exact same git commands, are not afraid to automate a push, and take good care of your .gitignore file, feel free to add something like this function to your ~/.bashrc and call it instead:

git_cord () {
  version=$(grep pyproject.toml -e '(?<=^version = ")(.*)(?=")' -Po)
  git add -u
  git commit -m "${version}"
  git tag "${version}"
  git push -u origin "${version}"
}

Getting started with gitlab-CI

The CI can in principle handle everything surrounding the deployment process, including version bumping and publishing. But the first requires that your CI can push to your repo (which has annoying side effects) and the latter that it can publish to your PyPI (which is risky, and makes debugging the CI a pain). I think it's not unusual to prefer to do those two steps by hand, so this minimal approach will only handle step 1 and 9. More extensive testing and build jobs can be included afterwards.

The correct setup of a CI depends on which one you plan to use. The list for github is long, so I'll instead focus on gitlab's builtin CI. It's free, has very little magic (which makes it comparably portable), and the binaries for the CI runners are open, free, and actually documented, so you can debug your CI locally or start and connect new runners if the free ones don't cut it for you.

Here is a small .gitlab-ci.yml that you can put into you project root in order to run the tests. Every single job in the pipeline (skipping setup and install commands) should also be executable in your dev environment, keeping it that way makes for a better maintainer-experience.

image: python:3.7-alpine

stages:
  - build
  - test

packaging:
  stage: build
  script:
    - pip install poetry
    - poetry build
  artifacts:
    paths: 
      - dist

pytest:
  stage: test
  script:
    - pip install dist/*.whl
    - pip install pytest
    - pytest

Setting up the build and test stage like this handles steps 1 and 9 in one swoop, while also running the test suite against the installed package instead of your source files. Though it will only work properly if you have have a src-layout in your project, which makes local sources unimportable from the project root. Some info on why that would be a good idea here and here.

Poetry can create a src-layout template you can move your code into with poetry new my_django_lib --src.

The changelog

While there are tools out there that automatically create a changelog from commit messages, keeping a good changelog is one of those things that benefit greatly from being cared for by hand. So, my advice is no automation for step 4.

One way to think about it is that the manual CHANGELOG file contains information that is relevant to your users, and should only feature information like new features, important bugfixes, and deprecations.

More fine grained information that might be important for contributors or plugin writers would be located in MRs, commit messages, or issue discussions, and should not make it into the CHANGELOG. You can try to collect it somehow, but navigating such an AUTOLOG is probably about as cumbersome as sifting through the primary sources I just mentioned.

So in short, the changelog-related parts of step 5 and 6 can be skipped.

---

CI with CD

Adding CD doesn't change too much, except that you don't have to release by hand any more. You can still release with poetry in case the CI is down, buggy, or you don't want to wait for the pipeline to release a hotfix.

This would alter the workflow in the following way:

• everyday work
  • write code (can't avoid this one yet)
  • document progress in commit messages and/or MRs (I prefer MRs, even for my own changes, and squash all commits on merge)
  • push to gitlab / merge MRs
• on release
  • create a tag, run poetry version and maybe poetry update
  • write release notes in CHANGELOG
  • push to gitlab

This addition to the former .gitlab-ci.yml file should work right away if you supply the secrets PYPI_USER and PYPI_PASSWORD:

stages:
  - build
  - test
  - release

[...]  # packaging and pytest unchanged

upload:
  stage: release
  only:
    - tags
    # Or alternatively "- /^v\d+\.\d+\.\d+/" if you also use non-release
    # tags, the regex only matches tags that look like this: "v1.12.0"
  script:
    - pip install poetry
    - poetry publish -u ${PYPI_USER} -p ${PYPI_PASSWORD} dist/*

---

Some useful links:

• .gitlab-ci.yml documentation
• list of predefined variables, this is where most of gitlab CI's obscurities lie
• the long version of my .gitlab-ci.yml template, with additional stages that may or may not be useful to you. It expects a src layout of your code.
  • lint: type checking, coverage, and code style
  • security: checking your own code and your dependencies for valnuarabilities
  • release.docs: public gitlab pages section where docs are served that are created automatically based on your docstrings
  • The build stage creates a wheelhouse from the poetry.lock file that can be used for installing dependencies later in favor of PyPI. This is a little faster, saves network bandwidth, and asserts the use of specific versions if you want to debug, but might be overkill and requires the use of a poetry pre-release.

---

^{[1] _{Among other things, poetry also 1) handles the virtualenv for you, 2) creates a hashed lockfile in case you need reproducible builds, and 3) makes contribution easier, since you only have to run "poetry install" after cloning a repo and are ready to go.}}