Property-Based Testing

From Examples to Properties

Example tests assert one input; property tests assert a rule across a generated space:

from hypothesis import given, strategies as st

def canonical(s):            # the function under test
    return "-".join(s.lower().split())

@given(st.text())
def test_idempotent(s):
    once = canonical(s)
    assert canonical(once) == once      # normalizing twice = normalizing once

@given(st.lists(st.integers(), min_size=1))
def test_sort_preserves(xs):
    assert sorted(xs) == sorted(reversed(xs))

Hypothesis runs many generated examples per call, and on failure it shrinks to a minimal counterexample (often a short string or small list) — the main advantage over hand-picked cases. Regressions auto-save: any failing example is replayed on every future run until the bug is fixed.

Building Strategies

from hypothesis import given, strategies as st, assume

@st.composite
def user(draw):
    name = draw(st.text(min_size=1, max_size=20))
    age = draw(st.integers(min_value=0, max_value=130))
    return {"name": name, "age": age}

@given(user())
def test_roundtrip(u):
    parsed = parse(serialize(u))
    assume(parsed is not None)     # discard invalid draws, don't assert-pass
    assert parsed == u

Composable builders cover most needs: st.one_of for unions, st.builds(Class, ...) for objects, st.sampled_from for enums, st.recursive for nested structures. Constrain with min_size/max_size rather than slicing results, and prefer assume for rare preconditions — but keep rejection rates low or Hypothesis reports a filtered_too_much health check.

Pitfalls and Interview Angles

• Tests must be pure. Reading the clock, network, or a shared DB makes properties flaky; inject fakes or seed the dependency.
• The database directory .hypothesis/ holds passing examples and failing regressions — keep it out of VCS or commit it deliberately so CI replays the same failures.
• Health checks. failed_health_check, too_slow, filter_too_much all mean the strategy or assumption logic is misdesigned, not that your code is correct.
• Stateful testing. RuleBasedStateMachine generates command sequences with @rule/@precondition for caches, queues, and DB layers where single-call properties are not enough.

Interview angle: name three cache properties (hit returns original value, size never exceeds bound, eviction keeps LRU order) and explain why a two-character shrunk counterexample teaches more than a 4 KB one.