gecrooks-python-template: Minimal viable setup for an open source, github hosted, python package
Installation for development
$ git clone https://github.com/gecrooks/gecrooks-python-template.git
$ cd gecrooks-python-template
$ pip install -e .[dev]
About: On the creation and crafting of a python project
This is a discussion of the steps needed to setup an open source, github hosted, python package ready for further development.
Naming
The first decision to make is the name of the project. And for python packages the most important criteria is that the name isn’t already taken on pypi, the repository from which we install python packages with pip. So we should do a quick Internet search: This name is available on pypi, there are no other repos of that name on github, and a google search doesn’t pull up anything relevant. So we’re good to go.
Note that github repo and pypi packages are named using dashes (-), but that the corresponding python module are named with underscores (_). (The reason for this dichotomy appears to be that underscores don’t work well in URLs, but dashes are frowned upon in filenames.)
License
The next decision is which of the plethora of Open Source licenses to use. We’ll use the Apache License, a perfectly reasonable, and increasingly popular choice.
(If you want to use a different license, replace the LICENSE file, update the License field in setup.cfg, and change the blurb at the top of each file of python code.)
Create repo
Next we need to initialize a git repo. It’s easiest to create the repo on github and clone to our local machine (This way we don’t have to mess around setting the origin and such like). Github will helpfully add a README.md, the license, and a python .gitignore for us. On Github, add a description, website url (typically pointing at readthedocs), project tags, and review the rest of github’s settings.
Note that MacOS likes to scatter .DS_Store folders around (they store the finder icon display options). We don’t want to accidentally add these to our repo. But this is a machine/developer issue, not a project issue. So if you’re on a mac you should configure git to ignore .DS_Store globally.
# specify a global exclusion list
git config --global core.excludesfile ~/.gitignore
# adding .DS_Store to that list
echo .DS_Store >> ~/.gitignore
Clone repo
On our local machine the first thing we do is create a new conda environment. (You have conda installed, right?) This way if we balls up the installation of some dependency (which happens distressingly often) we can nuke the environment and start again.
$ conda create --name GPT
$ source activate GPT
(GPT) $ python --version
Python 3.8.3
Now we clone the repo locally.
(GPT) $ git clone https://github.com/gecrooks/gecrooks-python-template.git
Cloning into 'gecrooks-python-template'...
remote: Enumerating objects: 4, done.
remote: Counting objects: 100% (4/4), done.
remote: Compressing objects: 100% (3/3), done.
remote: Total 4 (delta 0), reused 0 (delta 0), pack-reused 0
Unpacking objects: 100% (4/4), done.
(GPT) $ cd gecrooks-python-template
Lets tag this initial commit for posterities sake (And so I can link to the code at this instance).
(GPT) $ git tag v0.0.0
(GPT) $ git push origin v0.0.0
For reasons that are unclear to me the regular git push doesn’t push tags. We have push the tags explicitly by name. Note we need to specify a full MAJOR.MINOR.PATCH version number, and not just e.g. ‘0.1’, for technical reasons that have to do with how we’re going to manage package versions.
Branch
It’s always best to craft code in a branch, and then merge that code into the master branch.
$ git branch gec001-init
$ git checkout gec001-init
Switched to branch 'gec001-init'
I tend to name branches with my initials (so I know it’s my branch on multi-developer projects), a serial number (so I can keep track of the chronological order of branches), and a keyword (if I know ahead of time what the branch is for).
Packaging
Let’s complete the minimum viable python project. We need the actual python module, signaled by a (currently) blank __init__.py file.
(GPT) $ mkdir gecrooks_python_template
(GPT) $ touch gecrooks_python_template/__init__.py
Python standards for packaging and distribution seems to be in flux (again…). So following what I think the current standard is we need 3 files, setup.py, pyproject.toml, and setup.cfg.
The modern setup.py is just a husk:
#!/usr/bin/env python
import setuptools
if __name__ == "__main__":
setuptools.setup(use_scm_version=True)
Our only addition is use_scm_version=True, which activates versioning with git tags. More on that anon. Don’t forget to set executable permissions on the setup.py script.
$ chmod a+x setup.py
The pyproject.toml file (written in toml format) is a recent addition to the canon. It specifies the tools used to build the project.
# pyproject.toml
[build-system]
requires = ["setuptools>=42", "wheel", "setuptools_scm[toml]>=3.4"]
build-backend = "setuptools.build_meta"
# pyproject.toml
[tool.setuptools_scm]
Again, the parts with setuptools_scm are additions.
All of the rest of the metadata goes in setup.cfg (in INI format).
# Setup Configuration File
# https://docs.python.org/3/distutils/configfile.html
# [INI](https://docs.python.org/3/install/index.html#inst-config-syntax) file format.
[metadata]
# https://packaging.python.org/specifications/core-metadata/
# https://www.python.org/dev/peps/pep-0639/
# SPDX license short-form identifier, https://spdx.org/licenses/
Metadata-Version: 2.2
Name = gecrooks_python_template
Summary = Minimal viable setup for an open source, github hosted, python package
Long-Description = file:README.md
Long-Description-Content-Type = text/markdown
Keywords = python,template
Home-page = https://github.com/gecrooks/gecrooks-python-template/
Author = Gavin E. Crooks
Author-email = gec@threeplusone.com
License = Apache-2.0
License-File = LICENSE
# https://pypi.org/classifiers/
Classifiers=
Development Status :: 4 - Beta
Intended Audience :: Developers
Intended Audience :: Science/Research
Programming Language :: Python
Natural Language :: English
Operating System :: OS Independent
Programming Language :: Python :: 3
Programming Language :: Python :: 3.7
Programming Language :: Python :: 3.8
Programming Language :: Python :: 3.9
Topic :: Scientific/Engineering
Topic :: Software Development
Topic :: Software Development :: Libraries
Topic :: Software Development :: Libraries :: Python Modules
Typing :: Typed
[options]
zip_safe = True
python_requires = >= 3.7
packages = find:
install_requires =
importlib_metadata # required for python 3.7
numpy # example
setup_requires =
setuptools_scm
[options.extras_require]
dev =
pytest >= 4.6
pytest-cov
flake8
mypy
black
isort
sphinx
sphinxcontrib-bibtex
setuptools_scm
It’s good practice to support at least two consecutive versions of python. Starting with 3.9, python is moving to an annual release schedule. The initial 3.x.0 release will be in early October and the first bug patch 3.x.1 in early December, second in February, and so on. Since it takes many important packages some time to upgrade (e.g. numpy and tensorflow are often bottlenecks), one should probably plan to upgrade python support by February each year. Upgrading involves changing the python version numbers in the tests and config.cfg, and then cleaning up any __future__ or conditional imports, or other hacks added to maintain compatibility with older python releases.
We can now install our package (as editable -e, so that the code in our repo is live).
$ pip install -e .[dev]
The optional [dev] will install all of the extra packages we need for test and development, listed under [options.extras_require] above.
Versioning
Our project needs a version number (e.g. ‘3.1.4’). We’ll try and follow the semantic versioning conventions. But as long as the major version number is ‘0’ we’re allowed to break things.
There should be a single source of truth for this number. My favored approach is use git tags as the source of truth (Option 7 in the above linked list). We’re going to tag releases anyways, so if we also hard code the version number into the python code we’d violate the single source of truth principle. We use the setuptools_scm package to automatically construct a version number from the latest git tag during installation.
The convention is that the version number of a python packages should be available as packagename.__version__.
So we add the following code to gecrooks_python_template/config.py to extract the version number metadata.
try:
# python >= 3.8
from importlib import metadata as importlib_metadata # type: ignore
except ImportError: # pragma: no cover
# python == 3.7
import importlib_metadata # type: ignore # noqa: F401
__all__ = ["__version__", "about"]
package_name = "gecrooks_python_template"
try:
__version__ = importlib_metadata.version(package_name) # type: ignore
except Exception: # pragma: no cover
# package is not installed
__version__ = "?.?.?"
and then in gecrooks_python_template/__init__.py, we import this version number.
from .config import __version__ as __version__ # noqa: F401
We put the code to extract the version number in config.py and not __init__.py, because we don’t want to pollute our top level package namespace.
The various pragmas in the code above (“pragma: no cover” and “type: ignore”) are there because the conditional import needed for python 3.7 compatibility confuses both our type checker and code coverage tools.
about
One of my tricks is to add a function to print the versions of the core upstream dependencies. This can be extremely helpful when debugging configuration or system dependent bugs, particularly when running continuous integration tests.
# Configuration (> python -m gecrooks_python_template.about)
platform macOS-10.13.6-x86_64-i386-64bit
gecrooks-python-template 0.0.1
python 3.8.3
numpy 1.18.5
pytest 5.4.3
pytest-cov 2.10.0
flake8 3.8.3
mypy 0.780
sphinx 3.1.1
sphinxcontrib-bibtex 1.0.0
setuptools_scm 4.1.2
The about() function to print this information is placed in config.py. The file about.py contains the standard python command line interface (CLI),
if __name__ == '__main__':
import gecrooks_python_template
gecrooks_python_template.about()
It’s important that about.py isn’t imported by any other code in the package, else we’ll get multiple import warnings when we try to run the CLI.
Unit tests
Way back when I worked as a commercial programmer, the two most important things that I learned were source control and unit tests. Both were largely unknown in the academic world at the time.
(I was once talking to a chap who was developing a new experimental platform. The plan was to build several dozens of these gadgets, and sell them to other research groups so they didn’t have to build their own. A couple of grad students wandered in. They were working with one of the prototypes, and they’d found some minor bug. Oh yes, says the chap, who goes over to his computer, pulls up the relevant file, edits the code, and gives the students a new version of that file. He didn’t run any tests, because there were no tests. And there was no source control, so there was no record of the change he’d just made. That was it. The horror.)
Currently, the two main options for python unit tests appear to be unittest from the standard library and pytest. To me unittest feels very javonic. There’s a lot of boiler plate code and I believe it’s a direct descendant of an early java unit testing framework. Pytest, on the other hand, feels pythonic. In the basic case all we have to do is to write functions (whose names are prefixed with ‘test_’), within which we test code with asserts. Easy.
There’s two common ways to organize tests. Either we place tests in a separate directory, or they live in the main package along with the rest of the code. In the past I’ve used the former approach. It keeps the test organized and separate from the production code. But I’m going to try the second approach for this project. The advantage is that the unit tests for a piece of code live right next to the code being tested.
Let’s test that we can access the version number (There is no piece of code too trivial that it shouldn’t have a unit test.) In gecrooks_python_template/config_test.py we add
import gecrooks_python_template
def test_version():
assert gecrooks_python_template.__version__
and run our test. (The ‘python -m’ prefix isn’t strictly necessary, but it helps ensure that pytest is running under the correct copy of python.)
(GTP) $ python -m pytest
========================================================================================== test session starts ===========================================================================================
platform darwin -- Python 3.8.3, pytest-5.4.3, py-1.8.2, pluggy-0.13.1
rootdir: /Users/work/Work/Projects/gecrooks_python_template
collected 1 item
gecrooks_python_template/config_test.py . [100%]
=========================================================================================== 1 passed in 0.02s ============================================================================================
Note that in the main code we’ll access the package with relative imports, e.g.
from . import __version__
But in the test code we use absolute imports.
from gecrooks_python_template import __version__
In tests we want to access our code in the same way we would access it from the outside as an end user.
Test coverage
At a bare minimum the unit tests should run (almost) every line of code. If a line of code never runs, then how do you know it works at all?
So we want to monitor the test coverage. The pytest-cov plugin to pytest will do this for us. Configuration is placed in the setup.cfg file (Config can also be placed in a separate .coveragerc, but I think its better to avoid a proliferation of configuration files.)
# pytest configuration
[tool:pytest]
testpaths =
gecrooks_python_template
# Configuration for test coverage
#
# https://coverage.readthedocs.io/en/latest/config.html
#
# python -m pytest --cov
[coverage:paths]
source =
gecrooks_python_template
[coverage:run]
omit =
*_test.py
[coverage:report]
# Use ``# pragma: no cover`` to exclude specific lines
exclude_lines =
pragma: no cover
We have to explicitly omit the unit tests since we have placed the test files in the same directories as the code to test.
The pragam pragma: no cover is used to mark untestable lines. This often happens with conditional imports used for backwards compatibility between python versions.
Linting
We need to lint our code before pushing any commits. I like flake8. It’s faster than pylint, and I think better error messages. I will hereby declare:
The depth of the indentation shall be 4 spaces.
And 4 spaces shall be the depth of the indentation.
Two spaces thou shall not use.
And tabs are right out.
Four spaces is standard. Tabs are evil. I’ve worked on a project with 2-space indents, and I see the appeal, but I found it really weird.
Most of flake8’s defaults are perfectly reasonable and in line with PEP8 guidance. But even Linus agrees that the old standard of 80 columns of text is too restrictive. (Allegedly, 2-space indents was Google’s solution to the problem that 80 character lines are too short. Just make the indents smaller!) Raymond Hettinger suggests 90ish (without a hard cutoff), and black uses 88. So let’s try 88.
The configuration also lives in setup.cfg.
# flake8 linter configuration
[flake8]
max-line-length = 88
ignore = E203, W503
We need to override the linter on occasion. We add pragma such as # noqa: F401 to assert that no, really, in this case we do know what we’re doing.
Two other python code format tools to consider using are isort and black, The uncompromising code formatter. Isort sorts your import statements into a canonical order. And Black is the Model-T Ford of code formatting – any format you want, so long as it’s Black. I could quibble about some of Black’s code style, but in the end it’s just easier to blacken your code and accept black’s choices, and thereby gain a consistent coding style across developers.
The command make delint will run these isort and black on your code, with the right magic incantations so that they are compatible.
Copyright
It’s common practice to add a copyright and license notice to the top of every source file – something like this:
# Copyright 2019-, Gavin E. Crooks and contributors
#
# This source code is licensed under the Apache License, Version 2.0 found in
# the LICENSE.txt file in the root directory of this source tree.
I tend to forget to add these lines. So let’s add a unit test gecrooks_python_template/config_test.py::test_copyright to make sure we don’t.
def test_copyright():
"""Check that source code files contain a copyright line"""
exclude = set(['gecrooks_python_template/version.py'])
for fname in glob.glob('gecrooks_python_template/**/*.py', recursive=True):
if fname in exclude:
continue
print("Checking " + fname + " for copyright header")
with open(fname) as f:
for line in f.readlines():
if not line.strip():
continue
assert line.startswith('# Copyright')
break
API Documentation
Sphinx is the standard tool used to generate API documentation from the python source. Use the handy quick start tools.
$ mkdir docs
$ cd docs
$ sphinx-quickstart
The defaults are reasonable. Enter the project name and author when prompted.
Edit the conf.py, and add the following collection of extensions.
extensions = [
'sphinx.ext.autodoc',
'sphinx.ext.napoleon',
]
Autodoc automatically extracts documentation from docstrings, and napolean enables Google style python docstrings.
We also add a newline at the end of conf.py, since the lack of a blank line at the end upsets our linter.
Go ahead and give it a whirl. This won’t do anything interesting yet, but it’s a start.
$ make html
One problem is that sphinx creates three (initially) empty directories, _build, _static, and _templates. But we can’t add empty directories to git, since git only tracks files. The workaround is to add an empty .gitignore file to each of the _static and _templates directories. (Sphinx will create the _build directory when it needs it.)
$ touch _templates/.gitignore _build/.gitignore _static/.gitignore
$ git add -f _templates/.gitignore _build/.gitignore _static/.gitignore
$ git add Makefile *.*
# cd ..
Makefile
I like to add a Makefile with targets for all of the common development tools I need to run. This is partially for convenience, and partially as documentation, i.e. here are all the commands you need to run to test, lint, typecheck, and build the code (and so on.) I use a clever hack so that the makefile self documents.
(GTP) $ make
all Run all tests
test Run unittests
coverage Report test coverage
lint Lint check python source
delint Run isort and black to delint project
typecheck Static typechecking
docs Build documentation
docs-open Build documentation and open in webbrowser
docs-clean Clean documentation build
pragmas Report all pragmas in code
about Report versions of dependent packages
status git status -uno
build Setuptools build
clean Clean up after setuptools
The pragmas target searches the code and lists all of the pragmas that occur. Common uses of pragmas are to override the linter, tester, or typechecker. I also tend to scatter other keywords throughout my code: TODO (For things that need doing), FIXME (For code that’s broken, but I can’t fix right this moment), DOCME (code that needs more documentation), and TESTME (for code that needs more tests). In principle, production code shouldn’t have these pragmas. Either the problem should be fixed, or if it can’t be immediately fixed, it should become a github issue.
Readthedocs
We’ll host our API documentation on Read the Docs. We’ll need a basic configuration file, .readthedocs.yml.
version: 2
formats: []
sphinx:
configuration: docs/conf.py
python:
version: 3.8
I’ve already got a readthedocs account, so setting up a new project takes but a few minutes.
README.md
We add some basic information and installation instructions to README.mb. Github displays this file on your project home page (but under the file list, so if you have a lot of files at the top level of your project, people might not notice your README.)
A handy trick is to add Build Status and Documentation Status badges for Github actions tests and readthedocs. These will proudly declare that your tests are passing (hopefully). (See top of this file)
Continuous Integration
Another brilliant advance to software engineering practice is continuous integration (CI). The basic idea is that all code gets thoroughly tested before it’s added to the master branch.
Github now makes this very easy to setup with Github actions. They even provide basic templates. This testing workflow lives in .github/workflows/python-build.yml, and is a modification of Github’s python-package.yml workflow.
# This workflow will install Python dependencies, run tests and lint with a variety of Python versions
# For more information see: https://help.github.com/actions/language-and-framework-guides/using-python-with-github-actions
name: Python package
on:
push:
branches: [ master ]
pull_request:
branches: [ master ]
schedule:
- cron: "0 13 * * *" # Every day at 1pm UTC (6am PST)
jobs:
build:
runs-on: ubuntu-latest
strategy:
matrix:
python-version: ['3.7', '3.8']
steps:
- uses: actions/checkout@v2
- name: Set up Python $
uses: actions/setup-python@v2
with:
python-version: $
- name: Install dependencies
run: |
python -m pip install --upgrade pip
python -m pip install flake8 pytest
if [ -f requirements.txt ]; then pip install -r requirements.txt; fi
python -m pip install -e .[dev] # install package + test dependencies
- name: About
run: |
python -m $(python -Wi setup.py --name).about
- name: Lint with flake8
run: |
flake8 .
- name: Test with pytest
run: |
python -m pytest --cov-fail-under 100
- name: Typecheck with mypy
run: |
mypy
- name: Build documentation with sphinx
run: |
sphinx-build -M html docs docs/_build
Note that these tests are picky. Not only must the unit tests pass, but test coverage must be 100%, the code must be delinted, blackened, isorted, and properly typed, and the docs have to build without error.
It’s a good idea to set a cron job to run the test suite against the main branch on a regular basis (the schedule block above). This will alert you of problems caused by your dependencies updating. (For instance, one of my other projects just broke, apparently because flake8 updated it’s rules.)
Let’s add, commit, and push our changes.
$ git status
On branch gec001-init
Changes to be committed:
(use "git reset HEAD <file>..." to unstage)
new file: .readthedocs.yml
new file: .github/workflows/python-package.yml
new file: Makefile
modified: README.md
new file: docs/Makefile
new file: docs/_build/.gitignore
new file: docs/_static/.gitignore
new file: docs/_templates/.gitignore
new file: docs/conf.py
new file: docs/index.rst
new file: pyproject.toml
new file: gecrooks_python_template/__init__.py
new file: gecrooks_python_template/about.py
new file: gecrooks_python_template/config.py
new file: gecrooks_python_template/config_test.py
new file: setup.cfg
new file: setup.py
$ git commit -m "Minimum viable package"
...
$ git push --set-upstream origin gec001-init
...
If all goes well Github will see our push, and build and test the code in the branch. Probably all the tests won’t pass on the first try. It’s easy to forget something (which is why we have automatic tests). So tweak the code, and push another commit until the tests pass.
PyPi
We should now be ready to do a test submission to PyPI, The Python Package Index (PyPI). Follow the directions laid out in the python packaging documentation.
$ pip install -q wheel setuptools twine
...
$ git tag v0.1.0rc1
$ python setup.py sdist bdist_wheel
...
We tag our release candidate so that we get a clean version number (pypi will object to the development version numbers setuptools_scm generates if the tag or git repo isn’t up to date).
First we push to the pypi’s test repository.
(GTP) $ python -m twine upload --repository testpypi dist/*
You’ll need to create a pypi account if you don’t already have one.
Let’s make sure it worked by installing from pypi into a fresh conda environment.
(GTP) $ conda deactivate
$ conda create --name tmp
$ conda activate tmp
(tmp) $ pip install --index-url https://test.pypi.org/simple/ --no-deps gecrooks-python-template
(tmp) $ python -m gecrooks_python_template.about
(tmp) $ conda activate GTP
Merge and Tag
Over on github we create a pull request, wait for the github action checks to give us the green light once all the tests have passed, and then squash and merge.
The full developer sequence goes something like this
1.) Sync the master branch.
$ git checkout master
$ git pull origin master
(If we’re working on somebody else’s project, this step is a little more complicated. We fork the project on github, clone our fork to the local machine, and then set git’s ‘upstream’ to be the original repo. We then sync our local master branch with the upstream master branch
$ git checkout master
$ git fetch upstream
$ git merge upstream/master
This should go smoothly as long as you never commit directly to your local master branch.)
2.) Create a working branch.
$ git branch BRANCH
$ git checkout BRANCH
3.) Do a bunch of development on the branch, committing incremental changes as we go along.
4.) Sync the master branch with github (since other development may be ongoing.) (i.e. repeat step 1)
5.) Rebase our branch to master.
$ git checkout BRANCH
$ git rebase master
If there are conflicts, resolve them, and then go back to step 4.
6.) Sync our branch to github
$ git push
7.) Over on github, create a pull request to merge into the master branch
8.) Wait for the integration tests to pass. If they don’t, fix them, and then go back to step 4.
9.) Squash and merge into the master branch on github. Squashing merges all of our commits on the branch into a single commit to merge into the master branch. We generally don’t want to pollute the master repo history with lots of micro commits. (On multi-developer projects, code should be reviewed. Somebody other than the branch author approves the changes before the final merge into master.)
10.) Goto step 1. Back on our local machine, we resync master, create a new branch, and continue developing.
Tag and release
Assuming everything went well, you can now upload a release to pypi proper. We can add a github workflow to automatically upload new releases tagged on github. The only additional configuration is to upload PYPI_USERNAME and PYPI_PASSWORD to github as secrets (under you repo settings).
Conclusion
By my count we have 13 configuration files (In python, toml, yaml, INI, gitignore, Makefile, and plain text formats), 2 documentation files, one file of unit tests, and 3 files of code (containing 31 lines of code). We’re now ready to create a new git branch and start coding in earnest.
License
This software template is public domain. The included open-source software license LICENSE.txt and copyright lines are for illustrative purposes only. If you wish to use this template as the basis of your own project, you should feel free to assert your own copyrights (at the top of the python source code files) and substitute your own choice of software license.
Gavin E. Crooks (2020)
This is free and unencumbered software released into the public domain.
Anyone is free to copy, modify, publish, use, compile, sell, or
distribute this software, either in source code form or as a compiled
binary, for any purpose, commercial or non-commercial, and by any
means.
In jurisdictions that recognize copyright laws, the author or authors
of this software dedicate any and all copyright interest in the
software to the public domain. We make this dedication for the benefit
of the public at large and to the detriment of our heirs and
successors. We intend this dedication to be an overt act of
relinquishment in perpetuity of all present and future rights to this
software under copyright law.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
OTHER DEALINGS IN THE SOFTWARE.
