Continuous Integration#

Continuous integration (CI) is part of our development process and ensure that every piece of code or documentation which is contributed to SciPy is working and does not have unforeseen effects.

Note

Before submitting or updating your PR, please ensure that you tested your changes locally. See Checklist before submitting a PR and Running SciPy Tests Locally.

Workflows#

We run more than 20 different workflows with different versions of the dependencies, different architectures, etc. A PR must pass all these checks before it can be merged as to ensure a sustainable state of the project.

Apart from the unit tests, the documentation and examples in the docstrings are also checked. These are common failing workflows as Sphinx and doctests have very strict rules. These aspects are very important as documentation and examples are user facing elements. Ensures that these elements are properly rendered.

The logs can be long, but you will always find out why your build/test did not pass a check. Simply click on Details to access the logs.

Following is a list of all the different workflows in use. They are grouped by CI resources providers.

GitHub Actions#

  • Lint: PEP8 and code style

  • Windows Tests: test suite runs for Windows

  • Linux Tests: test suite runs for Linux

  • macOS Tests: test suite runs for macOS (x86_64)

  • Wheels builder: builds wheels for SciPy releases as well as nightly builds.

  • Check the rendered docs here!: live preview of the documentation

  • prerelease_deps_coverage_64bit_blas: use pre-released version of the dependencies and check coverage

  • gcc-9: build with minimal supported version of GCC, install the wheel, then run the test suite with python -OO

  • Array API: test Array API support

The test suite runs on GitHub Actions and other platforms cover a range of test/environment conditions: Python and NumPy versions (lowest-supported to nightly builds), 32-bit vs. 64-bit, different compilers, and more - for details, see the .yml configuration files.

CircleCI#

  • build_docs: build the documentation

  • build_scipy

  • run_benchmarks: verify how the changes impact performance

  • refguide_check: doctests from examples and benchmarks

CirrusCI#

  • Tests: test suite for specific architecture like musllinux, arm, aarch

  • Wheels: build and upload some wheels

Skipping#

Being an open-source project, we have access to a quota of CI resources. Ultimately, resources are limited and we should use them with care. This is why we ask you to verify your changes locally before pushing them.

Depending on the proposed change, you might want to skip part of the checks. It will be at the discretion of a maintainer to re-run some tests before integration.

Skipping CI can be achieved by adding a special text in the commit message:

  • [skip actions]: will skip GitHub Actions

  • [skip circle]: will skip CircleCI

  • [skip cirrus]: will skip CirrusCI

  • [docs only]: will skip all but the CircleCI checks and the linter

  • [lint only]: will skip all but the linter

  • [skip ci]: will skip all CI

Of course, you can combine these to skip multiple workflows.

This skip information should be placed on a new line. In this example, we just updated a .rst file in the documentation and ask to skip all but the relevant docs checks (skip Cirrus and GitHub Actions’ workflows):

DOC: improve QMCEngine examples.

[docs only]

Failures due to test duration#

Some CI jobs install pytest-fail-slow and report failures when the test execution time exceeds a threshold duration.

  • By default, all tests are subject to a 5 second limit; i.e., the option --fail-slow=5.0 is used in a “full” test job.

  • All tests not marked slow (@pytest.mark.slow) are subject to a 1 second limit; i.e. the option --fail-slow=1.0 is used in a “fast” test job.

  • Exceptions are made using the pytest.mark.fail_slow decorator; e.g. a test can be marked @pytest.mark.fail_slow(10) to give it a ten second limit regardless of whether it is part of the “fast” or “full” test suite.

If a test fails by exceeding the time limit at any point during the development of a PR, please adjust the test to ensure that it does not fail in the future. Even if new tests do not fail, please check the details of workflows that include “fail slow” in their name before PRs merge. These include lists of tests that are approaching (or have exceeded) their time limit. Due to variation in execution times, tests with execution times near the threshold should be adjusted to avoid failure even if their execution time were to increase by 50%; typical tests should have much greater margin (at least 400%). Adjustment options include:

  • Making the test faster.

  • Marking the test as slow, if it is acceptable to run the test on a reduced set of platforms.

  • Marking the test as xslow, if it is acceptable to run the test only occasionally.

  • Breaking up the test or parameterizing it, and possible marking parts of it as slow. Note that this does not reduce the total test duration, so other options are preferred.

  • For truly critical tests that are unavoidably slow, add an exception using pytest.mark.fail_slow.

See Running SciPy Tests Locally for more information about working with slow tests locally.

Wheel builds#

Wheels for SciPy releases and *nightly* builds are built using cibuildwheel in a Github Action. The Action runs:

  • when the commit message contains the text [wheel build]

  • on a scheduled basis once a week

  • when it is started manually.

  • when there is a push to the repository with a GitHub reference starting with refs/tags/v (and not ending with dev0)

The action does not run on forks of the main SciPy repository. The wheels that are created are available as artifacts associated with a successful run of the Action. When the Action runs on a schedule, or is manually started, the wheels are uploaded to the *scientific-python-nightly-wheels* repository.

It is not advised to use cibuildwheel to build scipy wheels on your own system as it will automatically install gfortran compilers and various other dependencies. Instead, one could use an isolated Docker container to build Linux wheels.