write-unit-test

Write Unit Test

Write a single-behavior, fast, isolated unit test using Arrange-Act-Assert that reads like documentation.

Why This Is Best Practice

Adopted by: Google, Microsoft, Stripe, and Amazon enforce single-behavior tests and the Arrange-Act-Assert (AAA) pattern in their internal testing guidelines. Google's Testing Blog and Google's "Software Engineering at Google" book (O'Reilly, 2020) codify AAA and DAMP as org-wide standards. Impact: Google's internal research (reported in "Software Engineering at Google") found that test suites following AAA with clear behavior names reduced debugging time after failures by ~40% because the failing test name and structure immediately identify the broken contract. Tests with multiple assertions per test hide failure root causes — engineers must re-read test body rather than reading the test name alone. Why best: DRY (Don't Repeat Yourself) in tests creates hidden coupling — a shared helper change silently breaks 30 tests. DAMP (Descriptive And Meaningful Phrases) tolerates duplication to keep each test self-contained and readable. This is the explicit recommendation in Jay Fields' "Working Effectively with Unit Tests" (2014) and in Google's testing guidelines over DRY-style shared setUp abuse.

Sources: Kent Beck, "Test-Driven Development: By Example" (2002); Google Testing Blog (testing.googleblog.com); Jay Fields, "Working Effectively with Unit Tests" (2014); "Software Engineering at Google" (Winters, Manshreck, Wright, O'Reilly 2020)

Steps

1. Identify one behavior

Name the single behavior to test before writing a line of code. The name is the test's contract:

<unit>_<condition>_<expectedOutcome>
# Examples:
Cart_addItem_incrementsQuantityWhenItemAlreadyPresent
PasswordValidator_validate_rejectsTooShortPasswords
PaymentService_charge_throwsOnExpiredCard

Refuse to write a test named test1 or testAll. If you cannot name the behavior, you do not know what to test yet.

2. Arrange — set up the world

Create exactly the state the test needs. Inline fixtures directly in the test (DAMP, not DRY). Do not use shared setUp for data that differs per test.

# Good — self-contained, readable
def test_cart_addItem_incrementsQuantityWhenItemAlreadyPresent():
    cart = Cart()
    cart.add(Item(id="sku-1", qty=1))
    # Act / Assert below ...

# Bad — reader must jump to setUp to understand the starting state
def setUp(self):
    self.cart = Cart()
    self.cart.add(Item(id="sku-1", qty=1))

DAMP vs DRY — where to draw the line

DAMP does not mean "never extract helpers." The rule is: inline data, extract behavior.

What to inline (DAMP)	What to extract (OK to DRY)
Input values, state, config specific to this test	make_user() / create_order() builder helpers that hide irrelevant noise
Expected output values	Shared assertion helpers (e.g., assert_valid_response(r))
Error messages and thresholds	Mock/fake setup for infrastructure (DB, HTTP client)

The 3-question test — before extracting shared test code, ask:

1. Does the extracted code contain a value that varies per test? → Keep inline.
2. Does a reader need to see this value to understand what the test proves? → Keep inline.
3. Is this purely mechanical plumbing that obscures the test's intent? → Extract.

# Inline — the specific discount rate IS the test's point
def test_cart_apply10PercentDiscount_reducesTotalByTen():
    cart = Cart(items=[Item(price=100)])
    cart.apply_discount(rate=0.10)          # 0.10 is the subject — must be visible
    assert cart.total() == 90.0

# Extract — building a valid User is noise; the test is about the email check
def test_user_changeEmail_rejectsInvalidFormat():
    user = make_user()                      # irrelevant details hidden
    with pytest.raises(ValueError):
        user.change_email("not-an-email")

def make_user(**overrides):
    defaults = {"name": "Alice", "role": "viewer", "email": "[email protected]"}
    return User(**{**defaults, **overrides})

Builder helpers are DAMP-safe when they expose relevant fields as overrides:

# Reader can see the field that matters — all else is noise
user = make_user(role="admin")
order = make_order(status="shipped", items=[Item(id="sku-1")])

Never share mutable fixtures. A helper that returns the same instance across tests causes hidden state leakage. Always return a fresh instance.

3. Act — invoke exactly one call

One call to the unit under test. If you need two calls to express the behavior, it is two behaviors — split the test.

    result = cart.add(Item(id="sku-1", qty=1))

4. Assert — verify one outcome

Assert one logical outcome (can be multiple assert calls if they verify one thing, e.g., status code + body). Never assert on implementation details (private state, internal method calls).

    assert cart.quantity("sku-1") == 2

5. Check speed and isolation

• Test must not touch the filesystem, network, database, clock, or random numbers without fakes/mocks.
• Test must run in < 100 ms. If slower, the unit is doing I/O — mock it.
• Tests must be order-independent: no shared mutable state between tests.

6. Review the test name as documentation

Read the test name aloud. A passing test should read as a true statement about the system. A failing test should tell a developer exactly what broke without opening the test body.

Rules

• One behavior per test — one act call, one logical assertion
• DAMP over DRY — repeat setup data inline rather than hiding it in helpers
• No logic in tests — no if, for, while inside test body
• Test names describe behavior, not implementation: rejectsExpiredCard not callsStripeRejectMethod
• Mocks for I/O (network, DB, clock); real objects for pure logic
• Never test private methods directly — test via the public API
• A test that never fails is not a test — verify it fails before the fix (red/green/refactor)

Examples

Good test

def test_passwordValidator_validate_rejectsTooShortPasswords():
    # Arrange
    validator = PasswordValidator(min_length=8)

    # Act
    result = validator.validate("short")

    # Assert
    assert result.is_valid is False
    assert "at least 8 characters" in result.error_message

Bad test (multiple behaviors, DRY abuse, implementation detail)

def test_passwordValidator():
    # Tests 3 things — which one broke?
    validator = self.validator  # hidden in setUp
    assert validator.validate("short").is_valid is False
    assert validator.validate("validpassword123").is_valid is True
    assert validator._rules["min_length"] == 8  # private detail

Common Mistakes

• Testing the mock — asserting that a mock was called rather than that the real outcome occurred. Tests coupling to implementation, not behavior.
• Giant arrange blocks — if setup exceeds 10 lines, the unit under test has too many dependencies. Refactor the unit, not the test.
• Asserting on everything — assert response.status == 200 and response.body == ... and log_called == True. Splits into three tests.
• Shared mutable fixtures — one test mutates shared state, the next test silently runs against corrupted state. Always create fresh instances.
• Ignoring flaky tests — a flaky test is a broken test. Mark it, file a bug, fix it within the sprint. Never skip-and-forget.