Prerequisite If you haven’t watched a presentation on property-based testing by John Hughes, please watch it – the presentation does a great job of explaining both: (i) what property-based testing is, and (ii) the power (nay, magic!) of property-based testing in finding hard-to-discover bugs.
Instead of enumerating expected input and output for unit tests, you write properties about your function that should hold true for all inputs. This lets you write concise, powerful tests.
If I had to summarize my understanding about property-based testing frameworks, it would be as such:
- Input data is randomly generated using composable and extensible generators.
- Output data is validated against invariant properties of the target function.
- A user-controlled knob can limit how extensive the testing needs to be, i.e., how many times to test the function with randomly generated data – hundreds of times, thousands of times, millions of times, etc.
- The framework is intelligent enough to generate simple test cases first, and gradually grow the complexity of input data as early tests pass.
- When an output violates the property invariant, the framework tries to shrink the input set to a minimal input set for which the invariant property is still violated. Often, the shrunk input will help directly pinpoint the root cause of invariant violation. When this works, one maybe reminded of Arthur C. Clarke’s famous “law”:
Any sufficiently advanced technology is indistinguishable from magic.
My Initial Take
I have never used a formal property-based testing framework before; though, I recall authoring unit tests that would validate results using randomly constructed input. The idea of using generated input seems fairly natural in the realm of testing and is likely used in many places in an ad hoc manner, i.e., without the use of composable generators. However, I’ve not authored any tests which attempt to shrink the input.
As a person made just learning about property-based testing, I have some doubts on the (i) efficacy vs. efficiency (“bang for the buck”) and (ii) wide-applicability of property-based (only) testing.
Firstly, on efficacy vs. efficiency: from developer perspective, classical unit testing (i.e., manually encoding output expectation for a given input) is easy to formulate (and grasp) as it directly validates the execution of a function. From computational perspective, classical unit testing also has the ability to hit key corner cases in a minimal number of tests; whereas, test cases based on randomly generated input data will reach corner cases in a computationally inefficient manner.
Secondly, and more importantly, it is not quite obvious to me if one could provide tightly-bound invariant properties for all functions needing test coverage. Tightly-bound invariants are required because loosely-bound invariants allow a faulty function to pass the tests. If we are unable to provide a set of tightly-bound invariants, property-based testing need to be augmented with output expectation unit tests.
Defining Tightly-bound Invariants
Connor Mendenhall presents an excellent introduction to Clojure’s property-based testing framework test.check in his blog article Check Your Work. If you have not heard of used test.check, you may gain a lot of value by reading that article before continuing on.
In the article, Connor also expresses his concerns the about wide-applicability of property-based testing. He describes his experience with using test.check framework to validate his solution to the Making Change “kata”:
After writing a few Simple-check [now renamed test.check] tests, I was stuck. I knew the specific case I was trying to describe—changing 80 cents with denominations of 25, 20, and 1 should return four 20 cent coins instead of three quarters and five pennies—but I was left scratching my head when I tried to generalize it.
I attempt to provide a solution to the problem of writing generalized test specs for the Making Change kata next.
A solution to the “Making Change” kata along with tightly-bound invariants for property-based tests is available on my GitHub.
I describe the key parts of the test coverage next. However, please refer to core_test.clj for complete implementation details.
Firstly, let’s cover the test spec to validate correctness of change, i.e., no loss of value:
1 2 3 4 5 6
correct-change test spec above,
gen-coinset is a generator that
yields a random, non-empty set of integers (“coins”) each time it is invoked.
coinset is one such instance. Similarly,
amount is an instance of a
randomly generated positive integer. The invariant states that the change
supplied by the
make-change function must add up to the original
Note that the
correct-change spec only makes sure the change amount is
correct, but it does not check minimality. Omitting either of the test
specs below would provide only a loose-bound on the invariants, allowing a
faulty function (for instance, one that simply returns
n 1-cent coins) to
pass the test.
Next, let’s discuss the test spec to validate the minimality of the simplest base case:
1 2 3 4 5
smallest-change test spec, the
gen-amount-from-coinset is a
generator that yields a random 2-tuple, consisting of: (i) a non-empty set of
coins, and (ii) an amount (that is equal in value to some coin in the set). For
this type of input, the invariant states that the minimal change should be
exactly one coin.
Next, we build a more “sophisticated” spec that exploits the base case along with an additional properly:
1 2 3 4 5 6 7 8 9 10 11
Combining the last two specs, using a coinset of 25-, 20-, and 1- cent coins (same as the one used by Connor in his article), we can provide change for 40-cents using two 20-cent coins. Similarly, we can provide change for 80-cents using four 20-cent coins. A greedy solution that uses two 25-cent coins, one 20-cent coin, and ten 1-cent coins (for a total of 13 coins) would fail the requirement for the target function to provide change for 80-cents using no more than 4 coins.
Many functions can be tightly-bound using a set of invariants. However, the key is to apply higher-order thinking than what we are used to with classical unit testing.
My concerns for wide-applicability of property-based tests are not entirely eliminated. Nevertheless, with a single short exercise, the universe of non-applicability has shrunk dramatically. I still need to use property-based tests in production code to get a better feel for their applicability and efficiency.
For anyone that is reading this, I would love to know if there is a reason that prevents the adoption of property-based (only) tests for a project. Please don’t hesitate to get in touch to discuss.