Hypothesis

Test faster, fix more

What happens when you run a test using Hypothesis? This article will help you understand.

The Python version of Hypothesis uses decorators to transform a test function that doesn’t use Hypothesis into one that does.

Consider the following example using py.test style testing:

from hypothesis.strategies import floats
from hypothesis import given


@given(floats(), floats())
def test_floats_are_commutative(x, y):
    assert x + y == y + x

The inner function takes two arguments, but the wrapping function defined by the @given decorator takes none and may be invoked as a normal test:

python -m pytest test_floats.py

And we see the following output from py.test:

    @given(floats(), floats())
    def test_floats_are_commutative(x, y):
>       assert x + y == y + x
E       assert (0.0 + nan) == (nan + 0.0)

test_floats.py:7: AssertionError

Falsifying example: test_floats_are_commutative(x=0.0, y=nan)

The test fails, because nan is a valid floating point number which is not equal to itself, and adding anything to nan yields nan.

When we ran this, Hypothesis invoked our test function with a number of randomly chosen values for the arguments until it found one that failed. It then attempted to shrink those values to produce a simpler one that would also fail.

If we wanted to see what it actually called our function with we can set the verbosity level. This can either be done in code with settings, or by specifying an environment variable:

from hypothesis.strategies import floats
from hypothesis import given, settings, Verbosity


@settings(verbosity=Verbosity.verbose)
@given(floats(), floats())
def test_floats_are_commutative(x, y):
    assert x + y == y + x
HYPOTHESIS_VERBOSITY_LEVEL=verbose python -m pytest test_floats.py

Any verbosity values explicitly passed in settings will override whatever is set at the environment level - the latter just provides a default.

Whichever one we choose, running it we’ll see output something like the following:


Trying example: test_floats_are_commutative(x=-0.05851890381391768, y=-6.060045836901702e+300)
Trying example: test_floats_are_commutative(x=-0.06323690311413645, y=2.0324087421708266e-308)
Trying example: test_floats_are_commutative(x=-0.05738038380011458, y=-1.5993500302384265e-308)
Trying example: test_floats_are_commutative(x=-0.06598754758697359, y=-1.1412902232349034e-308)
Trying example: test_floats_are_commutative(x=-0.06472919559855002, y=1.7429441378277974e+35)
Trying example: test_floats_are_commutative(x=-0.06537123121982172, y=-8.136220566134233e-156)
Trying example: test_floats_are_commutative(x=-0.06016703321602157, y=1.9718842567475311e-215)
Trying example: test_floats_are_commutative(x=-0.055257588875432875, y=1.578407827448836e-308)
Trying example: test_floats_are_commutative(x=-0.06313031758042666, y=1.6749023021600297e-175)
Trying example: test_floats_are_commutative(x=-0.05886897920547916, y=1.213699633272585e+292)
Trying example: test_floats_are_commutative(x=-12.0, y=-0.0)
Trying example: test_floats_are_commutative(x=4.0, y=1.7976931348623157e+308)
Trying example: test_floats_are_commutative(x=-9.0, y=0.0)
Trying example: test_floats_are_commutative(x=-38.0, y=1.7976931348623157e+308)
Trying example: test_floats_are_commutative(x=-24.0, y=1.5686642754811104e+289)
Trying example: test_floats_are_commutative(x=-10.0, y=nan)
Traceback (most recent call last):
  ...
AssertionError: assert (-10.0 + nan) == (nan + -10.0)

Trying example: test_floats_are_commutative(x=10.0, y=nan)
Traceback (most recent call last):
  ...
AssertionError: assert (10.0 + nan) == (nan + 10.0)

Trying example: test_floats_are_commutative(x=0.0, y=nan)
Traceback (most recent call last):
  ...
AssertionError: assert (0.0 + nan) == (nan + 0.0)

Trying example: test_floats_are_commutative(x=0.0, y=0.0)
Trying example: test_floats_are_commutative(x=0.0, y=inf)
Trying example: test_floats_are_commutative(x=0.0, y=-inf)
Falsifying example: test_floats_are_commutative(x=0.0, y=nan)

Notice how the first failing example we got was -10.0, nan but Hypothesis was able to turn that into 0.0, nan. That’s the shrinking at work. For a simple case like this it doesn’t matter so much, but as your examples get complicated it’s essential for making Hypothesis’s output easy to understand.

Trying example: test_floats_are_commutative(x=nan, y=0.0)
Traceback (most recent call last):
  ...
AssertionError: assert (nan + 0.0) == (0.0 + nan)

Trying example: test_floats_are_commutative(x=0.0, y=0.0)
Trying example: test_floats_are_commutative(x=inf, y=0.0)
Trying example: test_floats_are_commutative(x=-inf, y=0.0)
Falsifying example: test_floats_are_commutative(x=nan, y=0.0)

Now lets see what happens when we rerun the test:

Trying example: test_floats_are_commutative(x=0.0, y=nan)
Traceback (most recent call last):
  ...
AssertionError: assert (0.0 + nan) == (nan + 0.0)

Trying example: test_floats_are_commutative(x=0.0, y=0.0)
Trying example: test_floats_are_commutative(x=0.0, y=inf)
Trying example: test_floats_are_commutative(x=0.0, y=-inf)
Falsifying example: test_floats_are_commutative(x=0.0, y=nan)

Notice how the first example it tried was the failing example we had last time? That’s not an accident. Hypothesis has an example database where it saves failing examples. When it starts up it looks for any examples it has seen failing previously and tries them first before any random generation occurs. If any of them fail, we take that failure as our starting point and move straight to the shrinking phase without any generation.

The database format is safe to check in to version control if you like and will merge changes correctly out of the box, but it’s often clearer to specify the examples you want to run every time in the source code as follows:

from hypothesis.strategies import floats
from hypothesis import given, example


@example(0.0, float('nan'))
@given(floats(), floats())
def test_floats_are_commutative(x, y):
    assert x + y == y + x
Falsifying example: test_floats_are_commutative(x=0.0, y=nan)

If you run this in verbose mode it will print out Falsifying example: test_floats_are_commutative(x=0.0, y=nan) immediately and not try to do any shrinks. Values you pass in via example will not be shrunk. This is partly a technical limitation but it can often be useful as well.

Explicitly provided examples are run before any generated examples.

So, to recap and elaborate, when you use a test written using Hypothesis:

  1. Your test runner sees the decorated test as if it were a perfectly normal test function and invokes it.
  2. Hypothesis calls your test function with each explicitly provided @example. If one of these fails it stops immediately and bubbles up the exception for the test runner to handle.
  3. Hypothesis reads examples out of its database of previously failing examples. If any of them fail, it stops there and proceeds to the shrinking step with that example. Otherwise it continues to the generation step.
  4. Hypothesis tries generating a number of examples. If any of these raises an exception, it stops there and proceeds to the shrinking step. If none of them raise an exception, it silently returns and the test passes.
  5. Hypothesis takes the previously failing example it’s seen and tries to produce a “Simpler” version of it. Once it has found the simplest it possibly can, it saves that in the example database (in actual fact it saves every failing example in the example database as it shrinks, but the reasons why aren’t important right now).
  6. Hypothesis takes the simplest failing example and replays it, finally letting the test bubble up to the test runner.