glados 0.2.0

🍰 A property-based testing framework that tries to break your invariances.

Testing is tedious! At least that's what I thought before I stumbled over property-based testing – a simple approach that allows you to write fewer tests yet gain more confidence in your code. In traditional testing, you define concrete inputs and test whether they result in the desired output. In property based testing, you define certain conditions that are always true for any input (those are also called invariants). In mathematics, there's the ∀ operator for that. In Dart, now there's Glados.

Here are some benefits:

• ⚡ Write fewer tests. Let Glados figure out inputs that break your invariants.
• 🌌 Test for all possible inputs. Well, not literally all. But a huge variety.
• 🐜 Get a concise error report. Glados simplifies inputs that break your tests.
• 🤯 Understand the problem domain better by thinking of invariants.

Table of Contents

• Quickstart
• Comprehensive example
• How does it work?
• Customizing the exploration phase
• What's up with the name?
• Further info & resources

Quickstart

dev_dependencies:
  test: ...
  glados: ...

Use Glados<...>().test(...) instead of the traditional test(...).

// Running this test shows you that it fails for the input 21.
Glados<int>().test((a) {
  expect(a * 2, lessThan(42));
});

You can test with multiple inputs.

Glados2<String, int>().test((a, b) { ... });

Instead of using type parameters, you can customize inputs using any.

Glados(any.lowercaseLetter).test((letter) { ... });
Glados(any.nonEmptyList(any.positiveIntOrZero)).test((list) { ... });

You want to test with your own data classes? Use the @GenerateArbitrary() annotation and run pub run build_runner build! If that doesn't work, see this section.

part 'my_file.g.dart';

@GenerateArbitrary()
class MyClass { ... } // enums also work

Glados(any.myClass).test((a) { ... });

You can also customize the size of the generated inputs.

Comprehensive example

Suppose you write a function that tries to find the maximum in a list. I know – that's pretty basic – but it's enough to get you started. Here's an obviously wrong implementation:

/// If the list is empty, return null, otherwise the biggest item.
int max(List<int> input) => null;

To be sure that the function does the right thing, you might want to write some tests. Here's how those would look like in traditional unit testing:

test('maximum of empty list', () {
  expect(max([]), equals(null));
});
test('maximum of non-empty list', () {
  expect(max([40, 2, 10]), equals(40));
});

Not familiar with the syntax of the test package? Here are the docs.

Executing pub run test path/to/tests.dart should show that the second test fails.

In property-based testing, you look for invariants – conditions that should be true for any input. For example, if max produces null, the list should be empty:

Glados<List<int>>().test('maximum is only null if the list is empty', (list) {
  if (max(list) == null) {
    expect(list, isEmpty);
  }
});

Just create a Glados instance and call its test method instead of using the normal test function. Glados takes a generic type parameter – in this case, List<int>. It then tests your code with a variety of inputs of that type. All of them need to succeed for the whole test to succeed.

Running the test should produce something like this:

Tested 1 input, shrunk 25 times.
Failing for input: [0]
...

Glados discovered that a list with some content breaks the condition!

Let's modify our max function to pass this test:

int max(List<int> input) => 42;

We need to add another invariant to reject this function as well. Arguably the most obvious invariant for max is the following: The maximum should be greater than or equal to all items of the list:

Glados(any.nonEmptyList(any.int)).test('maximum is >= all items', (list) {
  var maximum = max(list);
  for (var item in list) {
    expect(maximum, greaterThanOrEqualTo(item));
  }
});

Instead of defining type parameters, you can also pass in arbitraries to Glados to customize which values are generated. You can find all available arbitraries as fields on the any value. In this case, we only test with non-empty lists because we handled the empty list in the first test.

Running the tests produces the following result:

Tested 35 inputs, shrunk 117 times.
Failing for input: [43]
...

Glados detected that the invariant breaks if the input list contains a 43.

Let's actually add a more reasonable implementation for max:

int max(List<int> input) {
  if (input.isEmpty) {
    return null;
  }
  var max = 0;
  for (var item in input) {
    if (item > max) {
      max = item;
    }
  }
  return max;
}

This fixes the tests, but still doesn't work for lists containing only negative values. So, let's add a final test:

Glados(any.nonEmptyList(any.int)).test('maximum is in the list', (list) {
  expect(list, contains(max(list)));
});

I'll leave implementing the function correctly to you, the reader.

But whatever solution you come up with, it'll be correct: Our tests aren't merely some arbitrary examples anymore. Rather, they correspond to the actual mathematical definition of max.

How does it work?

Glados works in two phases:

• 🌍 The exploration phase: Glados generates increasingly complex, random inputs until one breaks the invariant or the maximum number of runs is reached.
• 🐜 The shrinking phase: This phase only happens if Glados found an input that breaks the invariant. In this case, the input is gradually simplified and the smallest input that's still breaking the invariant is returned.

Arbitraries are responsible for generating and shrinking values. The Arbitrary class has two methods:

• T generate(Random random, int size) generates a new value of type T, using random as a source for randomness. The size argument is used as a rough estimate on how big or complex the returned value should be.
  For example, the arabitrary for int produces ints in the range from -size to size.
• Iterable<T> shrink(T input) takes a value and returns an Iterable containing similar, but smaller values. Smaller means that calling shrink repeatedly on the smaller values and their children etc., the program should eventually terminate (aka the transitive hull with regard to shrink should be finite and acyclic).

The basic types all have corresponding arbitraries implemented. All arbitraries can be found on any.

For most of the types, there are either arbitraries defined in other packages, or you'll be able to generate arbitraries using the @GenerateArbitrary() annotation (see Quickstart). Sometimes it makes sense to write new arbitraries manually.

For example, if you test code that expects email addresses, it may be inefficient to test the code with random Strings; if the tested code contains some sanity checks at the beginning, only a tiny fraction of values actually passes through the rest of the code.

In that case, create a custom arbitrary:

extension EmailAdressArbitrary on Any {
  Arbitrary<String> get emailAddress => arbitrary(
    generate: (random, size) => /* code for generating email addresses */,
    shrink: (emailAddress) => /* code for shrinking the given email address */,
  );
}

...

Glados(any.emailAddress).test('email test', (emailAddress) { ... });

You can also set an arbitrary as the default arbitrary for a type:

Any.setDefault<String>(any.emailAddress);

// Uses the any.emailAddress arbitrary based on the type parameters.
Glados<String>().test('blub', () { ... });

Customizing the exploration phase

You can also customize the exploration phase. To do that, you can use Explore, which is a configuration for certain values used during that phase.

For example, if you want to test some code with very big inputs, you might adjust Explore's parameters so that Glados starts with very big inputs and generates much bigger inputs after just a few runs:

Glados(any.email, Explore(
  initialSize: 100, // Start quite big
  speed: 10,        // and increase the input size by 10 each run,
  numRuns: 10,      // but only do 10 runs instead of 100.
)).test('my test', (input) {
  ...
});

Explore also has a random parameter, which you can provide with a custom Random instance. By default, Explore uses a Random instance created with a fixed seed so that your tests are deterministic.

What's up with the name?

GLaDOS is a very nice robot in the Portal game series. She's the head of the Aperture Science Laboratory facilities, where she spends the rest of her days testing. So I thought that's quite a fitting name. 🥔

By the way, both Portal games are great. If you haven't played them, definitely check them out.

Further info & resources

• Special thanks to @batteredgherkin for the Glados sticker.
• Here's the talk that got me into property-based testing.
• This article covers the topic in more detail.

🍰