pytest-benchmark 5.1.0

A ``pytest`` fixture for benchmarking code. It will group the tests into rounds that are calibrated to the chosen timer.

A pytest fixture for benchmarking code. It will group the tests into rounds that are calibrated to the chosen timer.

See calibration and FAQ.

• Free software: BSD 2-Clause License

Installation

pip install pytest-benchmark

Documentation

For latest release: pytest-benchmark.readthedocs.org/en/stable.

For master branch (may include documentation fixes): pytest-benchmark.readthedocs.io/en/latest.

Examples

But first, a prologue:

This plugin tightly integrates into pytest. To use this effectively you should know a thing or two about pytest first. Take a look at the introductory material or watch talks.

Few notes:

• This plugin benchmarks functions and only that. If you want to measure block of code or whole programs you will need to write a wrapper function.

• In a test you can only benchmark one function. If you want to benchmark many functions write more tests or use parametrization.

• To run the benchmarks you simply use pytest to run your “tests”. The plugin will automatically do the benchmarking and generate a result table. Run pytest --help for more details.

This plugin provides a benchmark fixture. This fixture is a callable object that will benchmark any function passed to it.

Example:

def something(duration=0.000001):
    """
    Function that needs some serious benchmarking.
    """
    time.sleep(duration)
    # You may return anything you want, like the result of a computation
    return 123

def test_my_stuff(benchmark):
    # benchmark something
    result = benchmark(something)

    # Extra code, to verify that the run completed correctly.
    # Sometimes you may want to check the result, fast functions
    # are no good if they return incorrect results :-)
    assert result == 123

You can also pass extra arguments:

def test_my_stuff(benchmark):
    benchmark(time.sleep, 0.02)

Or even keyword arguments:

def test_my_stuff(benchmark):
    benchmark(time.sleep, duration=0.02)

Another pattern seen in the wild, that is not recommended for micro-benchmarks (very fast code) but may be convenient:

def test_my_stuff(benchmark):
    @benchmark
    def something():  # unnecessary function call
        time.sleep(0.000001)

A better way is to just benchmark the final function:

def test_my_stuff(benchmark):
    benchmark(time.sleep, 0.000001)  # way more accurate results!

If you need to do fine control over how the benchmark is run (like a setup function, exact control of iterations and rounds) there’s a special mode - pedantic:

def my_special_setup():
    ...

def test_with_setup(benchmark):
    benchmark.pedantic(something, setup=my_special_setup, args=(1, 2, 3), kwargs={'foo': 'bar'}, iterations=10, rounds=100)

Screenshots

Normal run:

Compare mode (--benchmark-compare):

Histogram (--benchmark-histogram):

Also, it has nice tooltips.

Development

To run the all tests run:

tox

Credits

• Timing code and ideas taken from: https://github.com/vstinner/misc/blob/34d3128468e450dad15b6581af96a790f8bd58ce/python/benchmark.py

Changelog

5.1.0 (2024-10-30)

• Fixed broken hooks handling on pytest 8.1 or later (the TypeError: import_path() missing 1 required keyword-only argument: 'consider_namespace_packages' issue). Unfortunately this sets the minimum supported pytest version to 8.1.

5.0.1 (2024-10-30)

• Fixed bad fixture check that broke down then nbmake was enabled.

5.0.0 (2024-10-29)

• Dropped support for now EOL Python 3.8. Also moved tests suite to only test the latest pytest versions (8.3.x).

• Fix generate parametrize tests benchmark csv report errors (issue #268). Contributed by Johnny Huang in #269.

• Added the --benchmark-time-unit cli option for overriding the measurement unit used for display. Contributed by Tony Kuo in #257.

• Fixes spelling in some help texts. Contributed by Eugeniy in #267.

• Added new cprofile options:

  • --benchmark-cprofile-loops=LOOPS - previously profiling only ran the function once, this allow customization.

  • --benchmark-cprofile-top=COUNT - allows showing more rows.

  • --benchmark-cprofile-dump=[FILENAME-PREFIX] - allows saving to a file (that you can load in snakeviz, RunSnakeRun or other tools).

• Removed hidden dependency on py.path (replaced with pathlib).

4.0.0 (2022-10-26)

• Dropped support for legacy Pythons (2.7, 3.6 or older).

• Switched CI to GitHub Actions.

• Removed dependency on the py library (that was not properly specified as a dependency anyway).

• Fix skipping test in test_utils.py if appropriate VCS not available. Also fix typo. Contributed by Sam James in #211.

• Added support for pytest 7.2.0 by using pytest.hookimpl and pytest.hookspec to configure hooks. Contributed by Florian Bruhin in #224.

• Now no save is attempted if --benchmark-disable is used. Fixes #205. Contributed by Friedrich Delgado in #207.

3.4.1 (2021-04-17)

• Republished with updated changelog.

  I intended to publish a 3.3.0 release but I messed it up because bumpversion doesn’t work well with pre-commit apparently… thus 3.4.0 was set in by accident.

3.4.0 (2021-04-17)

• Disable progress indication unless --benchmark-verbose is used. Contributed by Dimitris Rozakis in #149.

• Added Python 3.9, dropped Python 3.5. Contributed by Miroslav Šedivý in #189.

• Changed the “cpu” data in the json output to include everything that cpuinfo outputs, for better or worse as cpuinfo 6.0 changed some fields. Users should now ensure they are an adequate cpuinfo package installed. MAY BE BACKWARDS INCOMPATIBLE

• Changed behavior of --benchmark-skip and --benchmark-only to apply early in the collection phase. This means skipped tests won’t make pytest run fixtures for said tests unnecessarily, but unfortunately this also means the skipping behavior will be applied to any tests that requires a “benchmark” fixture, regardless if it would come from pytest-benchmark or not. MAY BE BACKWARDS INCOMPATIBLE

• Added --benchmark-quiet - option to disable reporting and other information output.

• Squelched unnecessary warning when --benchmark-disable and save options are used. Fixes #199.

• PerformanceRegression exception no longer inherits pytest.UsageError (apparently a final class).

3.2.3 (2020-01-10)

• Fixed “already-imported” pytest warning. Contributed by Jonathan Simon Prates in #151.

• Fixed breakage that occurs when benchmark is disabled while using cprofile feature (by disabling cprofile too).

• Dropped Python 3.4 from the test suite and updated test deps.

• Fixed pytest_benchmark.utils.clonefunc to work on Python 3.8.

3.2.2 (2017-01-12)

• Added support for pytest items without funcargs. Fixes interoperability with other pytest plugins like pytest-flake8.

3.2.1 (2017-01-10)

• Updated changelog entries for 3.2.0. I made the release for pytest-cov on the same day and thought I updated the changelogs for both plugins. Alas, I only updated pytest-cov.

• Added missing version constraint change. Now pytest >= 3.8 is required (due to pytest 4.1 support).

• Fixed couple CI/test issues.

• Fixed broken pytest_benchmark.__version__.

3.2.0 (2017-01-07)

• Added support for simple trial x-axis histogram label. Contributed by Ken Crowell in #95).

• Added support for Pytest 3.3+, Contributed by Julien Nicoulaud in #103.

• Added support for Pytest 4.0. Contributed by Pablo Aguiar in #129 and #130.

• Added support for Pytest 4.1.

• Various formatting, spelling and documentation fixes. Contributed by Ken Crowell, Ofek Lev, Matthew Feickert, Jose Eduardo, Anton Lodder, Alexander Duryagin and Grygorii Iermolenko in #97, #105, #110, #111, #115, #123, #131 and #140.

• Fixed broken pytest_benchmark_update_machine_info hook. Contributed by Alex Ford in #109.

• Fixed bogus xdist warning when using --benchmark-disable. Contributed by Francesco Ballarin in #113.

• Added support for pathlib2. Contributed by Lincoln de Sousa in #114.

• Changed handling so you can use --benchmark-skip and --benchmark-only, with the later having priority. Contributed by Ofek Lev in #116.

• Fixed various CI/testing issues. Contributed by Stanislav Levin in #134, #136 and #138.

3.1.1 (2017-07-26)

• Fixed loading data from old json files (missing ops field, see #81).

• Fixed regression on broken SCM (see #82).

3.1.0 (2017-07-21)

• Added “operations per second” (ops field in Stats) metric – shows the call rate of code being tested. Contributed by Alexey Popravka in #78.

• Added a time field in commit_info. Contributed by “varac” in #71.

• Added a author_time field in commit_info. Contributed by “varac” in #75.

• Fixed the leaking of credentials by masking the URL printed when storing data to elasticsearch.

• Added a --benchmark-netrc option to use credentials from a netrc file when storing data to elasticsearch. Both contributed by Andre Bianchi in #73.

• Fixed docs on hooks. Contributed by Andre Bianchi in #74.

• Remove git and hg as system dependencies when guessing the project name.

3.1.0a2 (2017-03-27)

• machine_info now contains more detailed information about the CPU, in particular the exact model. Contributed by Antonio Cuni in #61.

• Added benchmark.extra_info, which you can use to save arbitrary stuff in the JSON. Contributed by Antonio Cuni in the same PR as above.

• Fix support for latest PyGal version (histograms). Contributed by Swen Kooij in #68.

• Added support for getting commit_info when not running in the root of the repository. Contributed by Vara Canero in #69.

• Added short form for --storage/--verbose options in CLI.

• Added an alternate pytest-benchmark CLI bin (in addition to py.test-benchmark) to match the madness in pytest.

• Fix some issues with --help in CLI.

• Improved git remote parsing (for commit_info in JSON outputs).

• Fixed default value for --benchmark-columns.

• Fixed comparison mode (loading was done too late).

• Remove the project name from the autosave name. This will get the old brief naming from 3.0 back.

3.1.0a1 (2016-10-29)

• Added --benchmark-columns command line option. It selects what columns are displayed in the result table. Contributed by Antonio Cuni in #34.

• Added support for grouping by specific test parametrization (--benchmark-group-by=param:NAME where NAME is your param name). Contributed by Antonio Cuni in #37.

• Added support for name or fullname in --benchmark-sort. Contributed by Antonio Cuni in #37.

• Changed signature for pytest_benchmark_generate_json hook to take 2 new arguments: machine_info and commit_info.

• Changed --benchmark-histogram to plot groups instead of name-matching runs.

• Changed --benchmark-histogram to plot exactly what you compared against. Now it’s 1:1 with the compare feature.

• Changed --benchmark-compare to allow globs. You can compare against all the previous runs now.

• Changed --benchmark-group-by to allow multiple values separated by comma. Example: --benchmark-group-by=param:foo,param:bar

• Added a command line tool to compare previous data: py.test-benchmark. It has two commands:

  • list - Lists all the available files.

  • compare - Displays result tables. Takes options:

    • --sort=COL

    • --group-by=LABEL

    • --columns=LABELS

    • --histogram=[FILENAME-PREFIX]

• Added --benchmark-cprofile that profiles last run of benchmarked function. Contributed by Petr Šebek.

• Changed --benchmark-storage so it now allows elasticsearch storage. It allows to store data to elasticsearch instead to json files. Contributed by Petr Šebek in #58.

3.0.0 (2015-11-08)

• Improved --help text for --benchmark-histogram, --benchmark-save and --benchmark-autosave.

• Benchmarks that raised exceptions during test now have special highlighting in result table (red background).

• Benchmarks that raised exceptions are not included in the saved data anymore (you can still get the old behavior back by implementing pytest_benchmark_generate_json in your conftest.py).

• The plugin will use pytest’s warning system for warnings. There are 2 categories: WBENCHMARK-C (compare mode issues) and WBENCHMARK-U (usage issues).

• The red warnings are only shown if --benchmark-verbose is used. They still will be always be shown in the pytest-warnings section.

• Using the benchmark fixture more than one time is disallowed (will raise exception).

• Not using the benchmark fixture (but requiring it) will issue a warning (WBENCHMARK-U1).

3.0.0rc1 (2015-10-25)

• Changed --benchmark-warmup to take optional value and automatically activate on PyPy (default value is auto). MAY BE BACKWARDS INCOMPATIBLE

• Removed the version check in compare mode (previously there was a warning if current version is lower than what’s in the file).

3.0.0b3 (2015-10-22)

• Changed how comparison is displayed in the result table. Now previous runs are shown as normal runs and names get a special suffix indicating the origin. Eg: “test_foobar (NOW)” or “test_foobar (0123)”.

• Fixed sorting in the result table. Now rows are sorted by the sort column, and then by name.

• Show the plugin version in the header section.

• Moved the display of default options in the header section.

3.0.0b2 (2015-10-17)

• Add a --benchmark-disable option. It’s automatically activated when xdist is on

• When xdist is on or statistics can’t be imported then --benchmark-disable is automatically activated (instead of --benchmark-skip). BACKWARDS INCOMPATIBLE

• Replace the deprecated __multicall__ with the new hookwrapper system.

• Improved description for --benchmark-max-time.

3.0.0b1 (2015-10-13)

• Tests are sorted alphabetically in the results table.

• Failing to import statistics doesn’t create hard failures anymore. Benchmarks are automatically skipped if import failure occurs. This would happen on Python 3.2 (or earlier Python 3).

3.0.0a4 (2015-10-08)

• Changed how failures to get commit info are handled: now they are soft failures. Previously it made the whole test suite fail, just because you didn’t have git/hg installed.

3.0.0a3 (2015-10-02)

• Added progress indication when computing stats.

3.0.0a2 (2015-09-30)

• Fixed accidental output capturing caused by capturemanager misuse.

3.0.0a1 (2015-09-13)

• Added JSON report saving (the --benchmark-json command line arguments). Based on initial work from Dave Collins in #8.

• Added benchmark data storage(the --benchmark-save and --benchmark-autosave command line arguments).

• Added comparison to previous runs (the --benchmark-compare command line argument).

• Added performance regression checks (the --benchmark-compare-fail command line argument).

• Added possibility to group by various parts of test name (the --benchmark-compare-group-by command line argument).

• Added historical plotting (the --benchmark-histogram command line argument).

• Added option to fine tune the calibration (the --benchmark-calibration-precision command line argument and calibration_precision marker option).

• Changed benchmark_weave to no longer be a context manager. Cleanup is performed automatically. BACKWARDS INCOMPATIBLE

• Added benchmark.weave method (alternative to benchmark_weave fixture).

• Added new hooks to allow customization:

  • pytest_benchmark_generate_machine_info(config)

  • pytest_benchmark_update_machine_info(config, info)

  • pytest_benchmark_generate_commit_info(config)

  • pytest_benchmark_update_commit_info(config, info)

  • pytest_benchmark_group_stats(config, benchmarks, group_by)

  • pytest_benchmark_generate_json(config, benchmarks, include_data)

  • pytest_benchmark_update_json(config, benchmarks, output_json)

  • pytest_benchmark_compare_machine_info(config, benchmarksession, machine_info, compared_benchmark)

• Changed the timing code to:

  • Tracers are automatically disabled when running the test function (like coverage tracers).

  • Fixed an issue with calibration code getting stuck.

• Added pedantic mode via benchmark.pedantic(). This mode disables calibration and allows a setup function.

2.5.0 (2015-06-20)

• Improved test suite a bit (not using cram anymore).

• Improved help text on the --benchmark-warmup option.

• Made warmup_iterations available as a marker argument (eg: @pytest.mark.benchmark(warmup_iterations=1234)).

• Fixed --benchmark-verbose’s printouts to work properly with output capturing.

• Changed how warmup iterations are computed (now number of total iterations is used, instead of just the rounds).

• Fixed a bug where calibration would run forever.

• Disabled red/green coloring (it was kinda random) when there’s a single test in the results table.

2.4.1 (2015-03-16)

• Fix regression, plugin was raising ValueError: no option named 'dist' when xdist wasn’t installed.

2.4.0 (2015-03-12)

• Add a benchmark_weave experimental fixture.

• Fix internal failures when xdist plugin is active.

• Automatically disable benchmarks if xdist is active.

2.3.0 (2014-12-27)

• Moved the warmup in the calibration phase. Solves issues with benchmarking on PyPy.

  Added a --benchmark-warmup-iterations option to fine-tune that.

2.2.0 (2014-12-26)

• Make the default rounds smaller (so that variance is more accurate).

• Show the defaults in the --help section.

2.1.0 (2014-12-20)

• Simplify the calibration code so that the round is smaller.

• Add diagnostic output for calibration code (--benchmark-verbose).

2.0.0 (2014-12-19)

• Replace the context-manager based API with a simple callback interface. BACKWARDS INCOMPATIBLE

• Implement timer calibration for precise measurements.

1.0.0 (2014-12-15)

• Use a precise default timer for PyPy.

? (?)

• README and styling fixes. Contributed by Marc Abramowitz in #4.

• Lots of wild changes.