Exploring unit-testing: unittest v pytest: FIGHT!

PyConUK 2014 talk by Tom Viner / @tomviner

Slides of conference talk

View slides full screen

(The original slides of my draft talk were a bit dry)

in the red corner we have: UNITTEST

  • this heavyweight fighter has been around for over 15 years!
  • was welcomed into the family, of the standard library, since Python 2.1
  • this is a fighter with heritage
    • unittest's nickname gives it away: PyUnit
    • we're talking xUnit lineage
    • I'm talking jUnit, sUnit aaaaaaallllllll the way back to Kent Beck

in the blue corner we have our challenger: PYTEST

  • this new kid on the block is here to make changes
  • trainer by Holger Krekel
  • pytest grew up, and broke away from the original py family
    • only the sentimental still call it py dot test


  • Years ago, another fighter was born into this battle
  • They call this one nosetests, and it came of an early pytest
  • The result is that they cover slightly different spaces now...
  • and so I have to announce that nose has been disqualified from this talk
    • Althought it's sometimes hard to stop these libraries making another appearance

3 rounds!

  • arrange

  • act

  • assert

  • 3 a's

  • 3 segments of testing
  • seconds out, ding diiiing!

round 1: arrange

  • getting things ready to act
  • (and clearing the mess up afterward)

  • prepare data

  • initalise objects
  • gentle mocking

per test Arrange using setUp method

class TestMyDataTable(unittest.TestCase):

    # Arrange
    def setUp(self):
        self.db_table = make_db_table()

    def test_table(self):
        # Act
        success = self.db_table.ACT()

        # Assert

    # I always end up doing your tidying up! >:-/
    def tearDown(self):
  • standard unittest
  • state is stored on self
  • rerun for every test method

Arrange using @fixture functions

# Arrange
def db_table():
    return make_db_table()

def test_table(db_table):
    # Act
    success = db_table.ACT()

    # Assert
    assert success
  • no setUp, no self, no class
  • (Django: no fixture file!)
  • variables are returned
  • test function argument connects
    • dependancy injection

per class Arrange using setUpClass

class TestMyDataTable(unittest.TestCase):
    def setUpClass(cls):
        cls.conn = get_db_connection()

    def setUp(self):
        self.db_table = self.conn.make_db_table()

    def test_one(self):

    def tearDown(self):

    def tearDownClass(cls):

So how would this run, with 3 test methods?

  • only make the db connection once
  • keep test isolation

cache @fixture functions using scope

(ignoring clean up)

def db_conn():
    return create_db_conn()

def db_table(db_conn):
    return db_conn.make_db_table()

def test_one(db_table):
  • chainable
  • cachble via scope argument
  • run once per:
    • function (default)
    • cls
    • module
    • session

what about making a todo list of clean ups as we went along?

don't shed a tearDown

class TestMyDataRow(unittest.TestCase):

    def setUp(self):
        self.db_table = make_db_table()
  • add multiple cleanups as you go
    • great for modular cleanup
  • no tearDown required

pytest fixture finalizers

def db_table(request):
    db_table = make_db_table()
    request.addfinalizer(db_table.drop) # can't pass args
    return db_table

neat alternative: the yield_fixture

def db_table():
    db_table = make_db_table()
    yield db_table
    # do finalising here:
  • like unittest's addCleanup
  • what about more complicated cleaning up?
  • neat alternative - like context manager
    • no tearDown and no addCleanup/addfinalizer

fixture have loads more features:

Paramatisation of fixtures

re-run tests with different resources

from databases import MySQL, Postgres
@pytest.fixture(params=[MySQL(), Postgres()])
def db(request):
    return request.param

round 1 over!

how did you score the two fighters on Arrange?

  • unittest: setUp
  • pytest: fixtures -- Advanced py.test Fixtures (Floris Bruynooghe) in CC1.4

round 2: act

  • time for a swiss intervention
  • as a half swiss, not everything has to be about battling, violence you know
  • round 2 act is neutral to which framework
  • so relax
  • call your live code,
  • some code you want to test (we call this the System Under Test btw)
  • that's all I can say about act

round 3: assert

  • out neutral zone now
  • straight into round 3!

round 3: assert

  • verify what happened
  • was exactly the right data returned?
  • did the correct side-effect occur?

assert: part i

happy path: asserting the expected truth

unittest assertMethods


core of basic asserts

self.assertEqual(x, y)
self.assertGreaterEqual(four, five)

per type - no need to remember

self.assertMultiLineEqual(s1, s2)
self.assertListEqual(l1, l2)

# make more with addTypeEqualityFunc(type, function


self.assertAlmostEqual(a, b)
self.assertRegexpMatches(text, regexp)
self.assertDictContainsSubset(subset, full)

pytest naked assert

assert result == expected
assert result2 != expected
assert x is not None
assert y >= z
assert dict1 == dict2

Note: - much simpler, no self.assertX methods to remember - have to implement things like assertAlmostEqual yourself - or use a plugin like (pytest-raisesregexp)[https://github.com/Walkman/pytest-raisesregexp]

nose makes an entry

nose.tools.assert_almost_equal(x, y)
  • pulled out of unittest
  • better together

pytest naked assert

def test_add():
    a = 1
    b = 2
    expected = 4
    assert add(a, b) + add(b, b) == expected

Test result

    def test_add():
        a = 1
        b = 2
        expected = 4
>       assert add(a, b) + add(b, b) == expected
E       assert (3 + 4) == 4
E        +  where 3 = add(1, 2)
E        +  and   4 = add(2, 2)

test_pytest_examples.py:12: AssertionError

assert: part ii

provoking failure:

ensuring exceptions raised

often we expect our code to raise exception in certain cases


class TestAdd(unittest.TestCase):
    def test_validation(self):
        with self.assertRaises(TypeError) as e:
            add('a', 1)

        self.assertRaisesRegexp(FooBarError, msg_re)


def test_validation():
    with pytest.raises(TypeError):
        add('a', 1)


  • arrange, act, assert
  • arrange: setUp v @fixture
  • act: please try the (framework) neutral swiss cheese
  • assert: assertThis v naked assert

  • monolithic class

    • with TestCase you must group your tests by common setUp
    • may prefer to group tests by the feature they test
  • with pytest you can group by related tests

    • reusable beyond a single class, possibly even share between projects
  • unittest is the standard

    • stood the test of time
  • pytest is the challenger
    • lots of functionality
    • rich ecosystem of plugins
  • in the end: you decide the winner!

Thank You

I've been @tomviner, any


link to slides - made with reveal.js

Did I mention we're hiring? See HogarthWW.com

we're a big Python shop, who do things like write webapps, have hackdays and host the London dojo

Comments !