Developer’s guide

The source code for MetaWards is available on GitHub.

The data needed to run a MetaWards simulation is also on GitHub in the MetaWardsData repository.

Setting up your computer

MetaWards requires Python >= 3.7, so please install this before continuing further. Anaconda provides easy installers for Python that work on a range of operating systems and that don’t need root or admin permissions.

Virtual environments

It is recommended that you develop MetaWards in a Python virtual environment. You can create a new environment in the directory venvs/metawards-devel by typing;

mkdir venvs
python -m venv venvs/metawards-devel

Feel free to place the environment in any directory you want.

Virtual environments provide sandboxes which make it easier to develop and test code. They also allow you to install Python modules without interfering with other Python installations.

You activate you environment by typing;

source venvs/metawards-devel/bin/activate

This will update your shell so that all python commands (such as python, pip etc.) will use the virtual environment. You can deactivate the environment and return to the “standard” Python using;

deactivate

If you no longer want the environment then you can remove it using

rm -rf venvs/metawards-devel

Developer dependencies

If is recommended that you have the following modules installed when developing metawards;

  • cython

  • numpy

  • flake8

  • pytest

  • sphinx (plus sphinx_issues sphinx_rtd_theme)

You can install these manually, or all at once using;

pip install -r requirements-dev.txt

Coding Style

MetaWards is written in Python 3 (>= 3.7), with time-critical functions written using Cython and parallelised using OpenMP. The Cython code is written strictly in comformant C, meaning that the package should compile and work on any system on which Python >= 3.7 runs. We ourselves are running production MetaWards models on ARM64 on Linux, and develop on X86-64 on Linux and Mac laptops. The program is tested with CI/CD on Windows 10, and we thank the windows users who’ve helped us make the tutorial cross-platform.

The aim of the Python port is to provide a simple and robust API that is a strong foundation for robust growth and scale-up of MetaWards, and one in which unnecessary implementation details are hidden from the user.

We aim as much as possible to follow a PEP8 python coding style and recommend that developers install and use a linter such as flake8.

For the Cython pyx code we also try to maintain a PEP8 style where possible, and recommend using a tool such as autopep8 to keep this style (it is used by the lead developers, so contributions will be formatted using it eventually ;-)).

We require that non-python code is strictly C. While C++ is an excellent language, it is too bulky for use in MetaWards and makes it more challenging to create portable binary distributions.

For ease of installation and support, we also minimise or bundle external dependencies (e.g. we use a bundled version of the binomial random number generator from numpy). This code has to run on a wide variety of architectures, operating systems and machines - some of which don’t have any graphic libraries, so please be careful when adding a dependency.

With this in mind, we use the following coding conventions:

Naming

We follow a Python style naming convention.

  • Packages: lowercase, singleword

  • Classes: CamelCase

  • Methods: snake_case

  • Functions: snake_case

  • Variables: snake_case

  • Source Files: snake_case with a leading underscore

Functions or variables that are private should be named with a leading underscore. This prevents them from being prominantly visible in Python’s help and tab completion.

Relative imports should be used at all times, with imports ideally delayed until they are needed.

For example, to import the Network object into a function that is in the utils module, you would type

from .._network import Network

network = Network()
network.run(...)

or to import the Parameters from code that is the main MetaWards package, you would type

from ._parameters import Parameters

parameters = Parameters()
parameters.add_seeds(filename="additional_seeds.dat")

Note that many classes are Python dataclasses, which are really useful for quick and safe development of code. Python dataclasses should be preferred over writing your own data-style classes.

Modules

MetaWards consists of the main module, e.g. metawards, plus a metawards.utils module that contains useful utilities.

The main module should be the focus of external developers, while the utils module should only be needed by developers of metawards itself.

In addition, there is a metawards.app module which contains the code for the various command-line applications (e.g. the metawards executable).

MetaWards uses a plugin-style interface for most new development. The plugins are for iterators, extractors, mixers and movers. Ideally, most new code should be added as one of these plugins. If you can’t fit your code into a plugin then please raise an issue to discuss your idea with the core developers, so that a way forward can be found. We really appreciate your help, and want to make sure that your ideas can be included in the most compatible way in the code.

To make MetaWards easy for new developers to understand, we have a set of rules that will ensure that only necessary public functions, classes and implementation details are exposed to the Python help system.

  • Module files containing implementation details are prefixed with an underscore, i.e. _parameters.py

  • Where possible, external packages are hidden inside each module, e.g. import sys as _sys

  • Each module file contains an __all__ variable that lists the specific items that should be imported.

  • The package __init__.py can be used to safely expose the required functionality to the user with:

from module import *

This results in a clean API and documentation, with all extraneous information, e.g. external modules, hidden from the user. This is important when working interactively, since IPython and Jupyter do not respect the __all__ variable when auto-completing, meaning that the user will see a full list of the available names when hitting tab. When following the conventions above, the user will only be able to access the exposed names. This greatly improves the clarity of the package, allowing a new user to quickly determine the available functionality. Any user wishing expose further implementation detail can, of course, type an underscore to show the hidden names when searching.

Workflow

Feature branches

First make sure that you are on the development branch of MetaWards:

git checkout devel

Now create and switch to a feature branch. This should be prefixed with feature, e.g.

git checkout -b feature-process

While working on your feature branch you won’t want to continually re-install in order to make the changes active. To avoid this, you can either make use of PYTHONPATH, e.g.

make
PYTHONPATH=./build/lib.{XXX} python script.py
PYTHONPATH=./build/lib.{XXX} pytest tests

(where {XXX} is the build directory for Cython on your computer, e.g. ./build/lib.macosx-10.9-x86_64-3.7 - remember that you need to type make to rebuild any Cython code and to copy your updated files into that directory)

or use the develop argument when running the setup.py script, i.e.

python setup.py develop

(this installs your current version of metawards into your current python environment)

Testing

When working on your feature it is important to write tests to ensure that it does what is expected and doesn’t break any existing functionality. Tests should be placed inside the tests directory, creating an appropriately named sub-directory for any new packages.

The test suite is intended to be run using pytest. When run, pytest searches for tests in all directories and files below the current directory, collects the tests together, then runs them. Pytest uses name matching to locate the tests. Valid names start or end with test, e.g.:

# Files:
test_file.py       file_test.py
# Functions:
def test_func():
   # code to perform tests...
   return

def func_test():
   # code to perform tests...
   return

We use the convention of test_* when naming files and functions.

Running tests

To run the full test suite, simply type:

pytest tests

To run tests for a specific sub-module, e.g.:

pytest tests/utils

To only run the unit tests in a particular file, e.g.:

pytest tests/test_integration.py

To run a specific unit tests in a particular file, e.g.:

pytest tests/test_read_variables.py::test_parameterset

To get more detailed information about each test, run pytests using the verbose flag, e.g.:

pytest -v

More details regarding how to invoke pytest can be found here.

Writing tests

Basics

Try to keep individual unit tests short and clear. Aim to test one thing, and test it well. Where possible, try to minimise the use of assert statements within a unit test. Since the test will return on the first failed assertion, additional contextual information may be lost.

Floating point comparisons

Make use of the approx function from the pytest package for performing floating point comparisons, e.g:

from pytest import approx

assert 0.1 + 0.2 == approx(0.3)

By default, the approx function compares the result using a relative tolerance of 1e-6. This can be changed by passing a keyword argument to the function, e.g:

assert 2 + 3 == approx(7, rel=2)
Skipping tests

If you are using test-driven development it might be desirable to write your tests before implementing the functionality, i.e. you are asserting what the output of a function should be, not how it should be implemented. In this case, you can make use of the pytest skip decorator to flag that a unit test should be skipped, e.g.:

@pytest.mark.skip(reason="Not yet implemented.")
def test_new_feature():
    # A unit test for an, as yet, unimplemented feature.
    ...
Parametrizing tests

Often it is desirable to run a test for a range of different input parameters. This can be achieved using the parametrize decorator, e.g.:

import pytest
from operator import mul

@pytest.mark.parametrize("x", [1, 2])
@pytest.mark.parametrize("y", [3, 4])
def test_mul(x, y):
    """ Test the mul function. """
    assert mul(x, y) == mul(y, x)

Here the function test_mul is parametrized with two parameters, x and y. By marking the test in this manner it will be executed using all possible parameter pairs (x, y), i.e. (1, 3), (1, 4), (2, 3), (2, 4).

Alternatively:

import pytest
from operator import sub
@pytest.mark.parametrize("x, y, expected",
                        [(1, 2, -1),
                         (7, 3,  4),
                         (21, 58, -37)])
def test_sub(x, y, expected):
    """ Test the sub function. """
    assert sub(x, y) == -sub(y, x) == expected

Here we are passing a list containing different parameter sets, with the names of the parameters matched against the arguments of the test function.

Testing exceptions

Pytest provides a way of testing your code for known exceptions. For example, suppose we had a function that raises an IndexError:

def indexError():
    """ A function that raises an IndexError. """
    a = []
    a[3]

We could then write a test to validate that the error is thrown as expected:

def test_indexError():
    with pytest.raises(IndexError):
        indexError()
Custom attributes

It’s possible to mark test functions with any attribute you like. For example:

@pytest.mark.slow
def test_slow_function():
    """ A unit test that takes a really long time. """
    ...

Here we have marked the test function with the attribute slow in order to indicate that it takes a while to run. From the command line it is possible to run or skip tests with a particular mark.

pytest mypkg -m "slow"        # only run the slow tests
pytest mypkg -m "not slow"    # skip the slow tests

The custom attribute can just be a label, as in this case, or could be your own function decorator.

Continuous integration and delivery

We use GitHub Actions to run a full continuous integration (CI) on all pull requests to devel and main, and all pushes to devel and main. We will not merge a pull request until all tests pass. We only accept pull requests to devel. We only allow pull requests from devel to main. In addition to CI, we also perform a build of the website on pushes to devel and tags to main. The website is versioned, so that old the docs for old versions of the code are always available. Finally, we have set up continuous delivery (CD) on pushes to main and devel, which build the pypi source and binary wheels for Windows, Linux (manylinux2010) and OS X. These are manually uploaded to pypi when we tag releases, but we expect to automate this process soon.

Documentation

MetaWards is fully documented using a combination of hand-written files (in the doc folder) and auto-generated api documentation created from NumPy style docstrings. See here for details. The documentation is automatically built using Sphinx whenever a commit is pushed to devel, which will then update this website.

To build the documentation locally you will first need to install some additional packages.

pip install sphinx sphinx_issues sphinx_rtd_theme

Then move to the doc directory and run:

make html

When finished, point your browser to build/html/index.html.

Committing

If you create new tests, please make sure that they pass locally before commiting. When happy, commit your changes, e.g.

git commit src/metawards/_new_feature.py tests/test_feature \
    -m "Implementation and test for new feature."

Remember that it is better to make small changes and commit frequently.

If your edits don’t change the MetaWards source code, or documentation, e.g. fixing typos, then please add ci skip to your commit message, e.g.

git commit -a -m "Updating docs [ci skip]"

This will avoid unnecessarily running the GitHub Actions, e.g. building a new MetaWards package, updating the website, etc. (the GitHub actions are configured in the file .github/workflows/main.yaml). To this end, we have provided a git hook that will append [ci skip] if the commit only modifies files in a blacklist that is specified in the file .ciignore (analagous to the .gitignore used to ignore untracked files). To enable the hook, simply copy it into the .git/hooks directory:

cp git_hooks/commit-msg .git/hooks

Any additional files or paths that shouldn’t trigger a re-build can be added to the .ciignore file.

Next, push your changes to the remote server, e.g.

# Push to the feature branch on the main MetaWards repo, if you have access.
git push origin feature

# Push to the feature branch your own fork.
git push fork feature

When the feature is complete, create a pull request on GitHub so that the changes can be merged back into the development branch. For information, see the documentation here.

Thanks

First, thanks to you for your interest in MetaWards and for reading this far. We hope you enjoy having a play with the code and having a go at adding new functionality, fixing bugs, writing docs etc.

We would also like to thank Lester Hedges and the BioSimSpace team who provided great advice to set up the above, and from whose GitHub repo most of the procedures, scripts and documentation above is derived.