Python Performance Testing with Austin
The pytest-austin plugin for pytest brings
Python performance testing right into your CI pipelines. It uses
Austin to profile your test runs without
any instrumentation. All you have to do is simply mark the tests on which you
want to execute checks for preventing performance regressions.
import pytest @pytest.mark.total_time(td(milliseconds=50), function="fibonacci") @pytest.mark.total_time("99%", line=8) @pytest.mark.total_time("50.3141592653%", line=9) def test_hello_default(): fibonacci(27) fibonacci(25)
Any failed tests will be reported by pytest at the end of the session. All the
collected statistics are written on a file prefixed with
austin_ and followed
by a truncated timestamp, inside pytest rootdir. You can drop it onto
Speedscope for a quick visual representation of your
================================================================ test session starts ================================================================ platform linux -- Python 3.6.9, pytest-6.0.1, py-1.9.0, pluggy-0.13.1 -- /home/gabriele/.cache/pypoetry/virtualenvs/pytest-austin-yu27Ep_e-py3.6/bin/python3.6 cachedir: .pytest_cache rootdir: /tmp/pytest-of-gabriele/pytest-226/test_austin_time_checks0 plugins: cov-2.10.0, austin-0.1.0 collecting ... collected 3 items test_austin_time_checks.py::test_lines PASSED test_austin_time_checks.py::test_check_fails PASSED test_austin_time_checks.py::test_check_succeeds PASSED =================================================================== Austin report =================================================================== austin 2.0.0 Collected stats written on /tmp/pytest-of-gabriele/pytest-226/test_austin_time_checks0/.austin_97148135487643.aprof 🕑 Sampling time (min/avg/max) : 376/3327/18019 μs 🐢 Long sampling rate : 87/87 (100.00 %) samples took longer than the sampling interval 💀 Error rate : 0/87 (0.00 %) invalid samples test_austin_time_checks.py::test_lines test_lines:19 (test_austin_time_checks.py) -16.0 ms (-78.2% of 20.5 ms) test_austin_time_checks.py::test_lines test_lines:18 (test_austin_time_checks.py) -4.1 ms (-10.1% of 40.3 ms) test_austin_time_checks.py::test_lines fibonacci (test_austin_time_checks.py) -9.3 ms (-18.6% of 50.0 ms) test_austin_time_checks.py::test_check_fails test_check_fails (test_austin_time_checks.py) +99.8 ms (9978.6% of 1000.0 μs) test_austin_time_checks.py::test_check_succeeds test_check_succeeds (test_austin_time_checks.py) -9.0 ms (-8.1% of 110.0 ms) ================================================================== 1 check failed =================================================================== ================================================================= 3 passed in 0.35s =================================================================
pytest-austin can be installed directly from PyPI with
pip install pytest-austin --upgrade
NOTE In order for the plugin to work, the Austin binary needs to be on the
PATH environment variable. See Austin
installation instructions to
see how you can easily install Austin on your platform.
For platform-specific issues and remarks, refer to the
Compatibility section below.
Once installed, the plugin will try to attach Austin to the pytest process in
order to sample it every time you run pytest. If you want to prevent Austin from
profiling your tests, you have to steal its mojo. You can do so with the
--steal-mojo command line argument.
The plugin looks for the
total_time marker on collected test items, which
takes a mandatory argument
time and three optional ones:
If you simply want to check that the duration of a test item doesn’t take longer
time, you can mark it with
can either be a
float (in seconds) or an instance of
from datetime import timedelta as td import pytest @pytest.mark.total_time(td(milliseconds=50)) def test_hello_default(): ...
In some cases, you would want to make sure that a function or method called on a
certain line in your test script executes in under a certain amount of time, say
5% of the total test time. You can achieve this like so
import pytest @pytest.mark.total_time("5%", line=7) def test_hello_default(): somefunction() fastfunction() # <- this is line no. 7 in the test script someotherfunction()
In many cases, however, one would want to test that a function or a method
called either directly or indirectly by a test doesn’t take more than a certain
overall time to run. This is where the remaining arguments of the
marker come into play. Suppose that you want to profile the procedure
that is called by method
foo of an object of type
Snafu. To ensure that
bar doesn’t take longer than, say, 50% of the overall test duration, you can
import pytest @pytest.mark.total_time("50%", function="bar") def test_snafu(): ... snafu = Snafu() ... snafu.foo() ...
You can use the
module argument to resolve function name clashes. For example,
if the definition of the function/method
bar occurs within the modules
someothermodule.py, but you are only interested in the
one defined in
somemodule.py, you can change the above into
import pytest @pytest.mark.total_time("50%", function="bar", module="somemodule.py") def test_snafu(): ... snafu = Snafu() ... snafu.foo() ...
And whilst you can also specify a line number, this is perhaps not very handy
and practical outside of test scripts themselves, unless the content of the
module is stable enough that line numbers don’t need to be updated very
When the pluing runs, it will produce an output containing lines of the form
test_austin_time_checks.py::test_lines test_lines:19 (test_austin_time_checks.py) -16.0 ms (-78.2% of 20.5 ms)
In this case, a negative number, such as
-16.0 ms, indicates that the total
time spent on
test_lines:19 (test_austin_time_checks.py) was 16.0 ms less
than the total allowed time specified with the
total_time marker, which in
this example is 20.5 ms. Hence, negative numbers indicate a successful check.
Failing tests will have a positive delta reported, e.g.
test_austin_time_checks.py::test_check_fails test_check_fails (test_austin_time_checks.py) +99.8 ms (9978.6% of 1000.0 μs)
This indicates that the total time spent on
test_check_fails (test_austin_time_checks.py) was 99.8 ms more than the
required threshold, which was set to 1 ms.
One can perform memory allocation checks with the
total_memory marker. The
first argument is
size, which can be a percentage of the total memory
allocation of the marked test case, as well as an absolute measure of the
maximum amount of memory, e.g.,
"24 MB". The
line are the same as for the
total_time marker. The extra
argument can be set to
True to check for the total net memory usage, that
is the difference between memory allocations and deallocations.
import pytest @pytest.mark.total_memory("24 MB") def test_snafu(): allota_memory()
In order to perform memory checks, you need to specify either the
all profile mode via the
The negative and positive memory deltas reported by the plugin in the report
behave like the time deltas described in the previous section. That is, a
negative memory delta indicates a successful check, whereas a positive delta
indicates a check that has failed.
When in the
all profile mode, you can perform both time and memory checks by
import pytest @pytest.mark.total_time(5.15) @pytest.mark.total_memory("24 MB") def test_snafu(): allota_memory_and_time()
If your tests spawn other Python processes, you can ask pytest-austin to profile
them too with the
--minime option. Note that if your tests are spawning too
many non-Python processes, the sampling rate might be affected because of the
way that Austin tries to discover Python child processes.
This plugins generate a report on terminal and dumps the collected profiling
statistics on the file system as well, for later analysis and visualisation. The
verbosity of the terminal report can be controlled with the
option. By default, it is set to
minimal, which means that only checks that
have failed will be reported. Use
full to see the results for all the checks
that have been detected and executed by the plugin.
Regarding the dump of the profiling statistics, the generated file is in the
Austin format by default (this is a generalisation of the collapsed stack
format). If you want the plugin to dump the data in either the
speedscope format, you can set the
--profile-format option accordingly.
This plugin has been tested on Linux, MacOS and Windows. Given that it relies on
Austin for sampling the frame stacks of
the pytest process, its compatibility considerations apply to pytest-austin as
On Linux, the use of
sudo is required, unless the
capability is granted to the Austin binary with, e.g.
sudo setcap cap_sys_ptrace+ep `which austin`
Then the use of
sudo is no longer required to allow Austin to attach and
On MacOS, the use of
sudo is also mandatory, unless the user that is
invoking pytest belongs to the
If you like pytest-austin and you find it useful, there are ways for you to
If you want to help with the development, then have a look at the open issues
and have a look at the contributing guidelines before you
open a pull request.