Minigun

A QuickCheck-like library for property-based unit-testing of Python programs.

Minigun is inspired by QCheck, which in turn was inspired by QuickCheck. Both are libraries that provide implementations for performing property-based unit-testing; for OCaml and Haskell respectively.

If you would like a bit of motivation as to why you should use a QuickCheck-like system for testing your project, then I would recommend that you watch:

• John Hughes - Testing the Hard Stuff and Staying Sane
• John Hughes - Certifying your car with Erlang

If you wish to learn more about the subject, I can recommend Jan Midtgaard's lecture materials; it is OCaml based but translates easily to other QuickCheck-like libraries for other languages.

Install

Minigun requires Python >=3.12. It is distributed with pip and can be installed with the following command:

pip install minigun-soren-n

Quick Start

Using the CLI (Recommended)

Create a test module in the tests/ directory:

# tests/my_tests.py
from minigun import prop, conj


@prop("reversing a list twice gives the original")
def test_reverse(lst: list[int]) -> bool:
    return list(reversed(list(reversed(lst)))) == lst


@prop("list length distributes over concatenation")
def test_length(xs: list[int], ys: list[int]) -> bool:
    return len(xs + ys) == len(xs) + len(ys)


spec = conj(test_reverse, test_length)

Run your tests with a time budget:

minigun --time-budget 30

Using as a Library

from minigun import prop, check


@prop("reversing a list twice gives the original")
def test_reverse(lst: list[int]) -> bool:
    return list(reversed(list(reversed(lst)))) == lst


if __name__ == "__main__":
    import sys

    sys.exit(0 if check(test_reverse) else 1)

Run directly:

python my_tests.py

Documentation

Full documentation and tutorials at Read The Docs.

Usage Guide

CLI Test Runner

# Run all tests in ./tests directory
minigun --time-budget 30

# Run tests from a different directory
minigun --time-budget 60 --test-dir my_tests

# Run specific test modules (only these are imported)
minigun --time-budget 45 --modules my_tests other_tests

# Run only the properties whose descriptions contain a substring
minigun --time-budget 30 --select "sensitivities"

# List available test modules
minigun --list-modules

# Quiet mode (for CI/CD)
minigun --time-budget 60 --output quiet

# JSON output (for automation)
minigun --time-budget 30 --output json

# Reproduce a failing run
minigun --time-budget 30 --seed 42

# Evaluate four properties at once in worker processes ('auto': one per CPU)
minigun --time-budget 30 --jobs 4

The CLI discovers Python files in the test directory that export a module-level spec: Spec. Every property in every module is evaluated and reported. Property descriptions identify properties across the whole run, so they must be unique across modules; a description used twice is reported as an error naming both modules before anything runs. The time budget is shared between properties in proportion to how many attempts their input domains are worth, and time a property leaves unspent flows to the properties after it. A small finite domain is worth enough attempts to draw every value with 99.9% confidence (11 for a bool, 88 for ten values); larger and unbounded domains share one cap. A property that declares its attempts always makes all of them (see Declaring Effort). Failures are reported with their shrunk counterexample in every output mode. A counterexample is announced as soon as it is found, before shrinking starts (on stderr in quiet and JSON output), and shrinking stops when the run's budget is spent: no further shrunk candidate is evaluated, and the smallest counterexample found so far is reported with a note that it may not be minimal.

Every run is seeded. When a property fails, the seed is printed so the exact run can be replayed with --seed, and each property draws from its own source derived from that seed, so a property's samples never depend on what else ran. To replay one failing property cheaply, pass the seed together with --select and part of its description: the property draws the same values as in the full run.

With --jobs N, N properties are evaluated at once, each in a worker process that imports the test modules itself. A property draws the same values whichever worker evaluates it, so a seed replays a parallel run too, and the report lists outcomes in the same order as a sequential run. The time budget is then worker time: N workers share out N times the remaining time, and no property runs past the end of the run. A law that scans or clears the fixture directories as a whole sees the fixtures of properties running beside it. Workers end with the run as an interpreter does: their atexit handlers run, so a process pool shared by a module's laws can be shut down there, and a child process still running after them is terminated.

Running Programmatically

from minigun.orchestrator import OutputMode, RunConfig, TestModule, run
from minigun.specify import prop


@prop("your property")
def my_property(x: int) -> bool:
    return x + 0 == x


if __name__ == "__main__":
    import sys

    config = RunConfig(time_budget=30.0, output=OutputMode.QUIET)
    sys.exit(0 if run(config, [TestModule("my_tests", my_property)]) else 1)

To run with several jobs, give each TestModule a load: a picklable callable, such as a module-level function, that returns the module's specification in a worker process, and set RunConfig(jobs=...). Workers are started with spawn, so the entry point must be guarded by if __name__ == "__main__":.

For structured outcomes without a reporter, drive minigun.specify.evaluate directly; the tutorial's "Running specifications programmatically" section shows how.

Writing Tests

Basic Properties

from minigun import prop


@prop("addition is commutative")
def test_add_commute(x: int, y: int) -> bool:
    return x + y == y + x

Custom Generators

from minigun import prop, context, generate as g


@context(g.int_range(1, 100), g.int_range(1, 100))
@prop("division reverses multiplication")
def test_div(x: int, y: int) -> bool:
    return (x * y) // y == x

Bounded Domains

import datetime
import math
import string

from minigun import prop, context, generate as g


@context(
    g.float_range(0.0, 0.25),
    g.dates(datetime.date(2000, 1, 1), datetime.date(2040, 12, 31)),
    g.bounded_strings(1, 5, string.ascii_uppercase),
)
@prop("discount factors lie in the unit interval")
def test_discount(rate: float, day: datetime.date, ticker: str) -> bool:
    years = (day - datetime.date(2000, 1, 1)).days / 365
    factor = math.exp(-rate * years)
    return 0.0 < factor <= 1.0 and ticker.isupper()

nonzero_int_range(lower, upper) draws integers other than zero, for divisors and step counts. float_range draws its bounds and, when the range holds zero, zero and tiny magnitudes down to the smallest subnormal float more often than a uniform draw would; pass edges=False for uniform draws only. datetimes(first, last, resolution) draws datetimes on a grid.

Preconditions

When a precondition relates several parameters, reject the arguments inside the law with assume (or discard()) instead of generating the parameters together:

from minigun import prop, context, assume, generate as g

side = g.int_range(1, 20)


@context(side, side, side)
@prop("proper triangles have a positive area")
def test_heron(a: int, b: int, c: int) -> bool:
    assume(a < b + c and b < a + c and c < a + b)
    return (a + b + c) * (-a + b + c) * (a - b + c) * (a + b - c) > 0

Rejected attempts count as discards, and shrunk counterexamples always satisfy the precondition.

Declaring Effort

A property whose attempts are expensive, or statistical, can declare how many attempts it is worth. check makes exactly that many, and so does a budgeted run, even when the property's share of the time runs out first: a declared count is what the property needs, not a limit to cut. Its declared attempts still weigh in the time shares of the other properties, so they leave it room, but a run whose declared properties overrun their share takes longer than its budget, and the summary shows by how much:

import random
from minigun import prop


@prop("the mean of uniform samples is close to one half", attempts=50)
def test_mean(rng: random.Random) -> bool:
    mean = sum(rng.random() for _ in range(1000)) / 1000
    return abs(mean - 0.5) < 0.1

Combining Properties

from minigun import prop, check, conj, neg


@prop("property 1")
def test_1(x: int) -> bool:
    return x + 0 == x


@prop("this law is false and a counterexample must be found")
def test_2(x: int) -> bool:
    return x * 2 == x


# Check both together; neg holds when its term is refuted
success = check(conj(test_1, neg(test_2)))

Randomness Inside a Law

Annotate a parameter with random.Random to receive a source that is reproducible from the run seed:

import random
from minigun import prop


@prop("shuffling preserves the elements")
def test_shuffle(xs: list[int], rng: random.Random) -> bool:
    shuffled = list(xs)
    rng.shuffle(shuffled)
    return sorted(shuffled) == sorted(xs)

FAQ

Q: What's a good time budget?

A: Start with 30-60 seconds for quick feedback. Use 2-5 minutes for thorough testing in CI/CD.

Q: How do I test larger input spaces?

A: Increase the time budget. Properties over unbounded domains absorb the extra time, up to 10000 attempts each per run.

Q: Can I customize test generation?

A: Yes, use the @context decorator with generators from minigun.generate, or write your own generator and shrinker. See the tutorial for details.

Q: Some properties are much more expensive than others. How do I balance them?

A: Declare their attempts with @prop("...", attempts=n). The time budget is shared in proportion to attempt limits, and time a property leaves unspent flows to the properties after it. A declared property always makes all its attempts, so the run takes as long as they need.

Q: A run seems to hang after a failure. What is it doing?

A: Shrinking: the counterexample found is being reduced by re-running the law on smaller arguments, and a smaller argument can be slower to evaluate than the original. The failure is announced, with its seed and original arguments, as soon as it is found (on stderr in quiet and JSON output), and shrinking stops when the time budget is spent. A single evaluation already running is not interrupted, so a law whose evaluations can take minutes can still overrun by one evaluation.

Q: How do I measure coverage?

A: Run the CLI under coverage: coverage run --source=mylib -m minigun --time-budget 60, then coverage report. To combine with a pytest suite, run both with coverage run -p and then coverage combine.

Q: How do I share helpers between test modules?

A: Put them in files whose names start with an underscore, or in a subdirectory, and import them relatively: from ._models import levy_models. Test modules are imported as a package of their directory.

Q: One test module is broken. Can I still run the others?

A: Yes: minigun --modules name ... imports only the selected modules.

Q: How do I reproduce a failing run?

A: Every failing run prints its seed. Pass it back with minigun --seed <n> (or check(spec, seed=n)) to replay the exact same generation. Add --select "<part of the description>" to replay only the failing property; it draws the same values as in the full run.

Q: My property passed but was it tested?

A: A property whose generators or assumptions discard most of its attempts (for example an over-restrictive g.filter or assume) fails with a message saying how many attempts were discarded, rather than passing silently.

Real-World Usage

The following projects use Minigun for testing:

• Minigun (self-testing)
• Tickle (parsing library)

If you have used Minigun for testing a public project, please file an issue with a link to add it to this list.