Testing Guidelines#

This section describes the pytest testing framework and format standards for tests in Astropy core, coordinated packages, and packages using the OpenAstronomy Packaging Guide. It also serves as recommendations for affiliated packages.

Testing Dependencies#

Most commonly, you should install the full suite of testing and development dependencies:

python -m pip install --editable '.[dev_all]'

This will provide all dependencies for running the full test suite using tox and pytest. It will also allow running tests via any IDE which supports pytest integration.

Running Tests#

There are two different ways to run Astropy tests: tox and pytest. Each of these invokes pytest to run the tests but each one addresses a different use-case.

tox#

The most robust way to run the tests (which can also be the slowest) is to make use of Tox, which is a general purpose tool for automating Python testing. One of the benefits of tox is that it first creates a source distribution of the package being tested, and installs it into a new virtual environment, along with any dependencies that are declared in the package, before running the tests. This can therefore catch issues related to undeclared package data, or missing dependencies. Since we use tox to run many of the tests on continuous integration services, it can also be used in many cases to reproduce issues seen on those services.

You can run the test suite with all optional dependencies with:

tox -e test-alldeps

Other useful invocations include:

tox -e test  # Run the tests with the minimal set of dependencies
tox -l -v  # Print a description of all available test environments
tox -e codestyle  # Run code style checks using ``ruff``

Note

It is suggested that you automate the code-style checks using the provided pre-commit hook, as described in the Pre-commit section.

You can pass options directly to pytest when running tox by adding a -- after the regular tox command. For example to enable verbose output and debugging use:

tox -e test -- -v --pdb

This can be used in conjunction with the -P option provided by the pytest-filter-subpackage plugin to run just part of the test suite.

Note that even though tox caches information, interactive debug and test sessions with tox can be quite slow. For this case, it may be better to set up a virtual environment with an editable install. Here, tox can still help by setting up a complete test environment, which one can then activate:

tox -e test-alldeps --develop --notest
source .tox/test-alldeps/bin/activate

Here, we use --notest to prevent tox from running the tests, since the idea is to do that oneself – using the pytest commands described below, targeting the relevant sub-package or test file.

pytest#

The test suite can also be run directly from the native pytest command, which is much faster than using tox for iterative development. This assumes you are working in an isolated development environment.

In the uncommon situation that one or more compiled extensions have changed, you will need to rebuild them by re-running the usual editable install command:

python -m pip install --editable '.[dev_all]'

It is possible to run only the tests for a particular subpackage or set of subpackages. For example, to run only the wcs and utils tests from the commandline:

pytest -P wcs,utils

You can also specify a single directory, a file (.py python or .rst doc file), or a specific test to check, rerun only tests that failed in the previous run, or require remote data:

pytest astropy/modeling
pytest astropy/wcs/tests/test_wcs.py
pytest astropy/units -k float_dtype_promotion
pytest astropy/units/tests/test_quantity.py::TestQuantityCreation::test_float_dtype_promotion
pytest astropy/wcs/index.rst
pytest --last-failed
pytest --remote-data=any

For more details, see the pytest invocation guide and the description of caching.

Test-running options#

Testing for open files#

The filterwarnings settings under [tool.pytest.ini_options] in the pyproject.toml file has an option which converts all unhandled warnings to errors during a test run. As a result, any open file(s) that throw ResourceWarning (except the specific ones already ignored) would fail the affected test(s).

Test coverage reports#

Coverage reports can be generated using the pytest-cov plugin (which is installed automatically when installing pytest-astropy) by using e.g.:

pytest --cov astropy --cov-report html

There is some configuration inside the pyproject.toml file that defines files to omit as well as lines to exclude.

Running tests in parallel#

It is possible to speed up astropy’s tests using the pytest-xdist plugin.

Once installed, tests can be run in parallel using the '-n' commandline option. For example, to use 4 processes:

pytest -n 4

Pass -n auto to create the same number of processes as cores on your machine.

Running tests on an installed astropy#

You can also run the tests on an installed version of astropy. First you need to ensure that the testing dependencies are installed:

python -m pip install "astropy[test]"

Note that you can include the --dry-run option to see what would be installed. In particular astropy itself should not be re-installed since it already exists. Then from any directory other than an astropy source repository, run the following:

pytest --pyargs astropy

You can also include other pytest options as needed.

Writing tests#

pytest has the following test discovery rules:

  • test_*.py or *_test.py files

  • Test prefixed classes (without an __init__ method)

  • test_ prefixed functions and methods

Consult the test discovery rules for detailed information on how to name files and tests so that they are automatically discovered by pytest.

Simple example#

The following example shows a simple function and a test to test this function:

def func(x):
    """Add one to the argument."""
    return x + 1

def test_answer():
    """Check the return value of func() for an example argument."""
    assert func(3) == 5

If we place this in a test.py file and then run:

pytest test.py

The result is:

============================= test session starts ==============================
python: platform darwin -- Python 3.x.x -- pytest-x.x.x
test object 1: /Users/username/tmp/test.py

test.py F

=================================== FAILURES ===================================
_________________________________ test_answer __________________________________

    def test_answer():
>       assert func(3) == 5
E       assert 4 == 5
E        +  where 4 = func(3)

test.py:5: AssertionError
=========================== 1 failed in 0.07 seconds ===========================

Where to put tests#

Package-specific tests#

Each package should include a suite of unit tests, covering as many of the public methods/functions as possible. These tests should be included inside each sub-package, e.g:

astropy/io/fits/tests/

tests directories should contain an __init__.py file so that the tests can be imported and so that they can use relative imports.

Interoperability tests#

Tests involving two or more sub-packages should be included in:

astropy/tests/

Regression tests#

Any time a bug is fixed, and wherever possible, one or more regression tests should be added to ensure that the bug is not introduced in future. Regression tests should include the ticket URL where the bug was reported.

Working with data files#

Tests that need to make use of a data file should use the ~astropy.utils.data.get_pkg_data_fileobj or ~astropy.utils.data.get_pkg_data_filename functions. These functions search locally first, and then on the astropy data server or an arbitrary URL, and return a file-like object or a local filename, respectively. They automatically cache the data locally if remote data is obtained, and from then on the local copy will be used transparently. See the next section for note specific to dealing with the cache in tests.

They also support the use of an MD5 hash to get a specific version of a data file. This hash can be obtained prior to submitting a file to the astropy data server by using the ~astropy.utils.data.compute_hash function on a local copy of the file.

Tests that may retrieve remote data should be marked with the @pytest.mark.remote_data decorator, or, if a doctest, flagged with the REMOTE_DATA flag. Tests marked in this way will be skipped by default by pytest to prevent test runs from taking too long. These tests can be run with pytest --remote-data=any.

It is possible to mark tests using @pytest.mark.remote_data(source='astropy'), which can be used to indicate that the only required data is from the http://data.astropy.org server. To enable just these tests, you can run the tests with pytest --remote-data=astropy.

For more information on the pytest-remotedata plugin, see pytest-remotedata.

Examples#

from ...config import get_data_filename

def test_1():
    """Test version using a local file."""
    #if filename.fits is a local file in the source distribution
    datafile = get_data_filename('filename.fits')
    # do the test

@pytest.mark.remote_data
def test_2():
    """Test version using a remote file."""
    #this is the hash for a particular version of a file stored on the
    #astropy data server.
    datafile = get_data_filename('hash/94935ac31d585f68041c08f87d1a19d4')
    # do the test

def doctest_example():
    """
    >>> datafile = get_data_filename('hash/94935')  # doctest: +REMOTE_DATA
    """
    pass

The get_remote_test_data will place the files in a temporary directory indicated by the tempfile module, so that the test files will eventually get removed by the system. In the long term, once test data files become too large, we will need to design a mechanism for removing test data immediately.

Tests that use the file cache#

By default, Astropy’s test configuration sets up a clean file cache in a temporary directory that is used only for that test run and then destroyed. This is to ensure consistency between test runs, as well as to not clutter users’ caches (i.e., the cache directory returned by ~astropy.config.get_cache_dir) with test files.

However, some test authors (especially for affiliated packages) may find it desirable to cache files downloaded during a test run in a more permanent location (e.g., for large data sets). To this end the ~astropy.config.set_temp_cache helper may be used. It can be used either as a context manager within a test to temporarily set the cache to a custom location, or as a decorator that takes effect for an entire test function (not including setup or teardown, which would have to be decorated separately).

Furthermore, it is possible to change the location of the cache directory for the duration of the test run via Environment variables.

Tests that create files#

Some tests involve writing files. These files should not be saved permanently. The pytest ‘tmp_path’ fixture allows for the convenient creation of temporary directories, which ensures test files will be cleaned up. Temporary directories can also be helpful in the case where the tests are run in an environment where pytest would otherwise not have write access.

Setting up/Tearing down tests#

In some cases, it can be useful to run a series of tests requiring something to be set up first. There are four ways to do this:

Module-level setup/teardown#

If the setup_module and teardown_module functions are specified in a file, they are called before and after all the tests in the file respectively. These functions take one argument, which is the module itself, which makes it very easy to set module-wide variables:

def setup_module(module):
    """Initialize the value of NUM."""
    module.NUM = 11

def add_num(x):
    """Add pre-defined NUM to the argument."""
    return x + NUM

def test_42():
    """Ensure that add_num() adds the correct NUM to its argument."""
    added = add_num(42)
    assert added == 53

We can use this for example to download a remote test data file and have all the functions in the file access it:

import os

def setup_module(module):
    """Store a copy of the remote test file."""
    module.DATAFILE = get_remote_test_data('94935ac31d585f68041c08f87d1a19d4')

def test():
    """Perform test using cached remote input file."""
    f = open(DATAFILE, 'rb')
    # do the test

def teardown_module(module):
    """Clean up remote test file copy."""
    os.remove(DATAFILE)

Class-level setup/teardown#

Tests can be organized into classes that have their own setup/teardown functions. In the following:

def add_nums(x, y):
    """Add two numbers."""
    return x + y

class TestAdd42:
    """Test for add_nums with y=42."""

    def setup_class(self):
        self.NUM = 42

    def test_1(self):
        """Test behavior for a specific input value."""
        added = add_nums(11, self.NUM)
        assert added == 53

    def test_2(self):
        """Test behavior for another input value."""
        added = add_nums(13, self.NUM)
        assert added == 55

    def teardown_class(self):
        pass

In the above example, the setup_class method is called first, then all the tests in the class, and finally the teardown_class is called.

Method-level setup/teardown#

There are cases where one might want setup and teardown methods to be run before and after each test. For this, use the setup_method and teardown_method methods:

def add_nums(x, y):
    """Add two numbers."""
    return x + y

class TestAdd42:
    """Test for add_nums with y=42."""

    def setup_method(self, method):
        self.NUM = 42

    def test_1(self):
    """Test behavior for a specific input value."""
        added = add_nums(11, self.NUM)
        assert added == 53

    def test_2(self):
    """Test behavior for another input value."""
        added = add_nums(13, self.NUM)
        assert added == 55

    def teardown_method(self, method):
        pass

Function-level setup/teardown#

Finally, one can use setup_function and teardown_function to define a setup/teardown mechanism to be run before and after each function in a module. These take one argument, which is the function being tested:

def setup_function(function):
    pass

def test_1(self):
   """First test."""
    # do test

def test_2(self):
    """Second test."""
    # do test

def teardown_function(function):
    pass

Property-based tests#

Property-based testing lets you focus on the parts of your test that matter, by making more general claims - “works for any two numbers” instead of “works for 1 + 2”. Imagine if random testing gave you minimal, non-flaky failing examples, and a clean way to describe even the most complicated data - that’s property-based testing!

pytest-astropy includes a dependency on Hypothesis, so installation is easy - you can just read the docs or work through the tutorial and start writing tests like:

from astropy.coordinates import SkyCoord
from hypothesis import given, strategies as st

@given(
    st.builds(SkyCoord, ra=st.floats(0, 360), dec=st.floats(-90, 90))
)
def test_coordinate_transform(coord):
    """Test that sky coord can be translated from ICRS to Galactic and back."""
    assert coord == coord.galactic.icrs  # floating-point precision alert!

Other properties that you could test include:

  • Round-tripping from image to sky coordinates and back should be lossless for distortion-free mappings, and otherwise always below 10^-5 px.

  • Take a moment in time, round-trip it through various frames, and check it hasn’t changed or lost precision. (or at least not by more than a nanosecond)

  • IO routines losslessly round-trip data that they are expected to handle

  • Optimised routines calculate the same result as unoptimised, within tolerances

This is a great way to start contributing to Astropy, and has already found bugs in time handling. See issue #9017 and pull request #9532 for details!

(and if you find Hypothesis useful in your research, please cite it!)

Parametrizing tests#

If you want to run a test several times for slightly different values, you can use pytest to avoid writing separate tests. For example, instead of writing:

def test1():
    assert type('a') == str

def test2():
    assert type('b') == str

def test3():
    assert type('c') == str

You can use the @pytest.mark.parametrize decorator to concisely create a test function for each input:

@pytest.mark.parametrize(('letter'), ['a', 'b', 'c'])
def test(letter):
    """Check that the input is a string."""
    assert type(letter) == str

As a guideline, use parametrize if you can enumerate all possible test cases and each failure would be a distinct issue, and Hypothesis when there are many possible inputs or you only want a single simple failure to be reported.

Tests requiring optional dependencies#

For tests that test functions or methods that require optional dependencies (e.g., Scipy), pytest should be instructed to skip the test if the dependencies are not present, as the astropy tests should succeed even if an optional dependency is not present. astropy provides a list of boolean flags that test whether optional dependencies are installed (at import time). For example, to load the corresponding flag for Scipy and mark a test to skip if Scipy is not present, use:

import pytest
from astropy.utils.compat.optional_deps import HAS_SCIPY

@pytest.mark.skipif(not HAS_SCIPY, reason='scipy is required')
def test_that_uses_scipy():
    ...

These variables should exist for all of Astropy’s optional dependencies; a complete list of supported flags can be found in astropy.utils.compat.optional_deps.

Any new optional dependencies should be added to that file, as well as to the relevant entries in the pyproject.toml file in the [project.optional-dependencies] section; typically, under all for dependencies used in user-facing code (e.g., h5py, which is used to write tables to HDF5 format), and in test_all for dependencies only used in tests (e.g., skyfield, which is used to cross-check the accuracy of coordinate transforms).

Testing warnings#

In order to test that warnings are triggered as expected in certain situations, pytest provides its own context manager pytest.warns that, completely analogously to pytest.raises (see below) allows to probe explicitly for specific warning classes and, through the optional match argument, messages. Note that when no warning of the specified type is triggered, this will make the test fail. When checking for optional, but not mandatory warnings, pytest.warns() can be used to catch and inspect them.

Note

With pytest there is also the option of using the recwarn function argument to test that warnings are triggered within the entire embedding function. This method has been found to be problematic in at least one case (pull request 1174).

Testing exceptions#

Just like the handling of warnings described above, tests that are designed to trigger certain errors should verify that an exception of the expected type is raised in the expected place. This is efficiently done by running the tested code inside the pytest.raises context manager. Its optional match argument allows to check the error message for any patterns using regex syntax. For example the matches pytest.raises(OSError, match=r'^No such file') and pytest.raises(OSError, match=r'or directory$') would be equivalent to assert str(err).startswith(No such file) and assert str(err).endswith(or directory), respectively, on the raised error message err. For matching multi-line messages you need to pass the (?s) flag to the underlying re.search, as in the example below:

with pytest.raises(fits.VerifyError, match=r'(?s)not upper.+ Illegal key') as excinfo:
    hdu.verify('fix+exception')
assert str(excinfo.value).count('Card') == 2

This invocation also illustrates how to get an ExceptionInfo object returned to perform additional diagnostics on the info.

Testing configuration parameters#

In order to ensure reproducibility of tests, all configuration items are reset to their default values when pytest starts up.

Sometimes you’ll want to test the behavior of code when a certain configuration item is set to a particular value. In that case, you can use the astropy.config.ConfigItem.set_temp context manager to temporarily set a configuration item to that value, test within that context, and have it automatically return to its original value.

For example:

def test_pprint():
    from ... import conf
    with conf.set_temp('max_lines', 6):
        # ...

Marking blocks of code to exclude from coverage#

Blocks of code may be ignored by the coverage testing by adding a comment containing the phrase pragma: no cover to the start of the block:

if this_rarely_happens:  # pragma: no cover
    this_call_is_ignored()

Image tests with pytest-mpl#

Running image tests#

We make use of the pytest-mpl plugin to write tests where we can compare the output of plotting commands with reference files on a pixel-by-pixel basis (this is used for instance in astropy.visualization.wcsaxes). We use the hybrid mode with hashes and images.

To run the Astropy tests with the image comparison, use e.g.:

tox -e py311-test-image-mpl380-cov

However, note that the output can be sensitive to the operating system and specific version of libraries such as freetype. In general, using tox will result in the version of freetype being pinned, but the hashes will only be correct when running the tests on Linux. Therefore, if using another operating system, we do not recommend running the image tests locally and instead it is best to rely on these running in an controlled continuous integration environment.

Writing image tests#

The README.rst for the plugin contains information on writing tests with this plugin. Once you have added a test, and push this to a pull request, you will likely start seeing a test failure because the figure hash is missing from the hash libraries (see the next section for how to proceed).

Rather than use the @pytest.mark.mpl_image_compare decorator directly, you should make use of the @figure_test convenience decorator which sets the default tolerance and style to be consistent across the astropy core package, and also automatically enables access to remote data:

from astropy.tests.figures import figure_test

@figure_test
def test_figure():
    fig, ax = plt.subplots()
    ...
    return fig

You can optionally pass keyword arguments to @figure_test and these will be passed on to mpl_image_compare:

@figure_test(savefig_kwargs={'bbox_inches': 'tight'})
def test_figure():
    ...

Failing tests#

When existing tests start failing, it is usually either because of a change in astropy itself, or a change in Matplotlib. New tests will also fail if you have not yet updated the hash library.

In all cases, you can view a webpage with all the existing figures where you can check whether any of the figures are now wrong, or if all is well. The link to the page for each tox environment that has been run will be provided in the list of statuses for pull requests, and can also be found in the CircleCI logs. If any changes/additions look good, you can download from the summary page a JSON file with the hashes which you can use to replace the existing one in astropy/tests/figures.

New hash libraries#

When adding a new tox environment for image testing, such as for a new Matplotlib or Python version, the tests will fail as the hash library does not exist yet. To generate it, you should run the tests the first time with:

tox -e <envname> -- --mpl-generate-hash-library=astropy/tests/figures/<envname>.json

for example:

tox -e py311-test-image-mpl380-cov -- --mpl-generate-hash-library=astropy/tests/figures/py311-test-image-mpl380-cov.json

Then add and commit the new JSON file and try running the tests again. The tests may fail in the continuous integration if e.g. the freetype version does not match or if you generated the JSON file on a Mac or Windows machine - if that is the case, follow the instructions in Failing tests to update the hashes.

As an alternative to generating the JSON file above, you can also simply copy a previous version of the JSON file and update any failing hashes as described in Failing tests.

Generating reference images#

You do not need to generate reference images for new tests or updated reference images for changed tests - when pull requests are merged, a CircleCI job will automatically update the reference images in the astropy-figure-tests repository.

Writing doctests#

A doctest in Python is a special kind of test that is embedded in a function, class, or module’s docstring, or in the narrative Sphinx documentation, and is formatted to look like a Python interactive session–that is, they show lines of Python code entered at a >>> prompt followed by the output that would be expected (if any) when running that code in an interactive session.

The idea is to write usage examples in docstrings that users can enter verbatim and check their output against the expected output to confirm that they are using the interface properly.

Furthermore, Python includes a doctest module that can detect these doctests and execute them as part of a project’s automated test suite. This way we can automatically ensure that all doctest-like examples in our docstrings are correct.

The Astropy test suite automatically detects and runs any doctests in the astropy source code or documentation, or in packages using the Astropy test running framework. For example doctests and detailed documentation on how to write them, see the full doctest documentation.

For more information on the pytest-doctestplus plugin used by Astropy, see pytest-doctestplus.

Skipping doctests#

Sometimes it is necessary to write examples that look like doctests but that are not actually executable verbatim. An example may depend on some external conditions being fulfilled, for example. In these cases there are a few ways to skip a doctest:

  1. Next to the example add a comment like: # doctest: +SKIP. For example:

    >>> import os
>>> os.listdir('.')  # doctest: +SKIP

    In the above example we want to direct the user to run os.listdir('.') but we don’t want that line to be executed as part of the doctest.

    To skip tests that require fetching remote data, use the REMOTE_DATA flag instead. This way they can be turned on using the --remote-data flag when running the tests:

    >>> datafile = get_data_filename('hash/94935')  # doctest: +REMOTE_DATA

  2. Astropy’s test framework adds support for a special __doctest_skip__ variable that can be placed at the module level of any module to list functions, classes, and methods in that module whose doctests should not be run. That is, if it doesn’t make sense to run a function’s example usage as a doctest, the entire function can be skipped in the doctest collection phase.

    The value of __doctest_skip__ should be a list of wildcard patterns for all functions/classes whose doctests should be skipped. For example:

    __doctest_skip__ = ['myfunction', 'MyClass', 'MyClass.*']

    skips the doctests in a function called myfunction, the doctest for a class called MyClass, and all methods of MyClass.

    Module docstrings may contain doctests as well. To skip the module-level doctests include the string '.' in __doctest_skip__.

    To skip all doctests in a module:

    __doctest_skip__ = ['*']

  3. In the Sphinx documentation, a doctest section can be skipped by making it part of a doctest-skip directive:

    .. doctest-skip::

    >>> # This is a doctest that will appear in the documentation,
    >>> # but will not be executed by the testing framework.
    >>> 1 / 0  # Divide by zero, ouch!

    It is also possible to skip all doctests below a certain line using a doctest-skip-all comment. Note the lack of :: at the end of the line here:

    .. doctest-skip-all

All doctests below here are skipped...

  4. __doctest_requires__ is a way to list dependencies for specific doctests. It should be a dictionary mapping wildcard patterns (in the same format as __doctest_skip__) to a list of one or more modules that should be importable in order for the tests to run. For example, if some tests require the scipy module to work they will be skipped unless import scipy is possible. It is also possible to use a tuple of wildcard patterns as a key in this dict:

    __doctest_requires__ = {('func1', 'func2'): ['scipy']}

    Having this module-level variable will require scipy to be importable in order to run the doctests for functions func1 and func2 in that module.

    In the Sphinx documentation, a doctest requirement can be notated with the doctest-requires directive:

    .. doctest-requires:: scipy

    >>> import scipy
    >>> scipy.hamming(...)

Skipping output#

One of the important aspects of writing doctests is that the example output can be accurately compared to the actual output produced when running the test.

The doctest system compares the actual output to the example output verbatim by default, but this not always feasible. For example the example output may contain the __repr__ of an object which displays its id (which will change on each run), or a test that expects an exception may output a traceback.

The simplest way to generalize the example output is to use the ellipses .... For example:

>>> 1 / 0
Traceback (most recent call last):
...
ZeroDivisionError: integer division or modulo by zero

This doctest expects an exception with a traceback, but the text of the traceback is skipped in the example output–only the first and last lines of the output are checked. See the doctest documentation for more examples of skipping output.

Ignoring all output#

Another possibility for ignoring output is to use the # doctest: +IGNORE_OUTPUT flag. This allows a doctest to execute (and check that the code executes without errors), but allows the entire output to be ignored in cases where we don’t care what the output is. This differs from using ellipses in that we can still provide complete example output, just without the test checking that it is exactly right. For example:

>>> print('Hello world')
We don't really care what the output is as long as there were no errors...

Handling float output#

Some doctests may produce output that contains string representations of floating point values. Floating point representations are often not exact and contain roundoffs in their least significant digits. Depending on the platform the tests are being run on (different Python versions, different OS, etc.) the exact number of digits shown can differ. Because doctests work by comparing strings this can cause such tests to fail.

To address this issue, the pytest-doctestplus plugin provides support for a FLOAT_CMP flag that can be used with doctests. For example:

>>> 1.0 / 3.0  # doctest: +FLOAT_CMP
0.333333333333333311

When this flag is used, the expected and actual outputs are both parsed to find any floating point values in the strings. Those are then converted to actual Python float objects and compared numerically. This means that small differences in representation of roundoff digits will be ignored by the doctest. The values are otherwise compared exactly, so more significant (albeit possibly small) differences will still be caught by these tests.

Continuous integration#

Overview#

Astropy uses the following continuous integration (CI) services:

  • GitHub Actions for Linux, OS X, and Windows setups (Note: GitHub Actions does not have “allowed failures” yet, so you might see a fail job reported for your PR with “(Allowed Failure)” in its name. Still, some failures might be real and related to your changes, so check it anyway!)

  • CircleCI for visualization tests

These continuously test the package for each commit and pull request that is pushed to GitHub to notice when something breaks.

In some cases, you may see failures on continuous integration services that you do not see locally, for example because the operating system is different, or because the failure happens with only 32-bit Python.

Maintainers have the option to run comparative benchmark using GitHub Actions to test a new pull request against the current main branch. It uses the benchmarks from astropy-benchmarks. It is important to note that these benchmarks can be flaky as they run on virtual machines (and thus shared hardware) but they should give a general idea of the performance impact of a pull request.