You should consider testing for the impossible.

Cartoon image of a stick figure tester considering whether or not to test for black swans believing they don't exist while a black swan is standing behind them

The other day, a friend of mine asked me a question about a discussion they were having at his place of work. It regarded some code where he had implemented some correction to invalid data, and had decided to put unit tests to check that the correction happened. He had received feedback to remove the unit tests, on the grounds that the condition was impossible.

Do you think unit tests should test current or correct behavior?

John

I.E. there is a combination of inputs that in practice never happens. Ideally the code should assert that is the case, but it does not and we are adding new test coverage. Should there be a unit test asserting the behavior of the code in the incorrect case or should the behavior be left undefined by the test suite.

This is not an easy question to answer. I am going to use a situation similar to what my friend described to consider different ways of looking at this. Here are the main points regarding the unit in question and the conditions under which it operates

  1. The unit receives as input instructions to process items
  2. There are two types of items the code processes, CheckEntity and CheckAudit
  3. There are different categories of operation, add, update, execute_check
  4. The system is built such that type==CheckEntity AND category==execute_check is expected to never happen
  5. If there were an instance that type==CheckEntity AND category==execute_check did occur, it creates an ambiguous data state that confuses further processing

An example input might look like this:

{
  "category": "execute_check",
  "data": CheckAudit {
    ...
  }
}

Lets pretend the unit under test is a static method called Processor.processOperation(). The question is whether there ought to be a unit test that does the following:

// the pattern in this test is vague and wouldn't run because we
// haven't discussed the actual implementation and expected behavior
// in the article yet, but the main point is the test is checking a
// negative condition that is believed to never happen prior to the
// call
public void processOperation_executecheck_checkentity_handlesnegativecase()
{
  var checkEntity = new CheckEntity( ...whatever we use to initialize this...);
  var operation = { category: "execute_check", data: checkEntity};
   Assert.AreEqual( -whatever we expect-, Processor.processOperation(operation));
}

I am using a static method because it makes for shorter examples. Given the type of system I am imagining here, I would probably consider a static method and how it suggests the presence of a singleton with a big nasty chain of dependencies behind it a bad design choice.

Let’s consider some of the possibilities.

What is the test checking for?

Unit tests almost always perform confirmatory checks against some expected behavior. We could require several behaviors from the unit under test in this case.

  1. Do nothing
  2. Do nothing the caller would notice, but emit some event to the logs to track later if it does happen
  3. Throw an exception or abort further operation
  4. Return an error code
  5. Transform the data to mitigate confusion and errors in further processing

For the first two, there is no expected behavior from the unit in response to that condition. One could write a unit test with the errant data condition as an input, and then assert that nothing happens than if it were not errant. In the case of writing to logs, one could assert the entry happened. The thing is, there is no contract that any consumer of the API is counting on, no behavior they rely on being protected or enforced, so in that regard the unit test offers nothing for them. A unit test in this case is not a bad thing, but other than demonstrating coverage, it isn’t guarding much.

In the case of the last three, all of those decisions create a behavior that both the caller and downstream consumers of the code will rely on. In the case of the throw or the error code, the caller will need to handle both of those, and downstream processes are relying on never seeing an operation with this errant state. In the case of transformation, the caller will not need to handle anything, and downstream consumers will need to know to handle whatever that transformation looks like, but likewise should not ever expect the errant data state.

Let’s look at what the unit test might look like if the code will throw (this varies depending on the testing framework you use and the language the code is written in):

[TestMethod]
[Expect Exception(typeof InvalidOperationException)]
public void processOperation_executecheck_checkentity_handlesnegativecase()
{
  var checkEntity = new CheckEntity( ...whatever we use to initialize this...);
  var operation = { category: "execute_check", data: checkEntity};
  Processor.processOperation(operation);
}

The unit test above will help protect the contract, especially for downstream processes.

Use unit tests to detect violation of contracts that other code will depend on. How the code handles invalid conditions is part of that contract.

But it’s impossible, so why bother?

It might seem unnecessary to add unit tests for a state that is impossible for the system to get into. If the upstream processes of our processOperation() call never produce a combination of execute_check and CheckEntity, why bother with the test?

Checking for the supposedly impossible or not feels like one of those cases which separate testers and developers. The developers say “That cannot happen, why bother?”, and then the tester says “Oh yeah… let me show you…” and goes in search of a way to make it happen. Or the tester runs the test below the layer where the impossible state is guarded, and then the developer objects to the entire premise.

We should avoid thinking of it in that way. It makes for melodrama that might amuse us sometimes, but pragmatically this is a matter of whether or not we are using engineering practices that help us know when something went wrong. The question is whether we are being pragmatic or not.

Because the state is impossible. Right?

Is it really though?

We might believe something to be impossible, and it really might be, but that is usually only because of the current state of the system. Systems change. Maybe more importantly, what we believe about the system and the truth about the system may be very different. There are some things which make knowing what is impossible or not easier or more difficult.

Built in language and platform enforcement

Sometimes the language we use to write the code, or the technology we build it on, make certain kinds of states or situations impossible. For example, imagine the two following methods, written in two different languages:

// written in C#, with strongly typed inputs
public int countMyFingers(int hands)

// written in JavaScript, with loosely typed inputs
function countMyFingers(hands)

In the first version of the method, passing in a non-integer value for hands is an uninteresting and meaningless test because the language and the compiler make doing so impossible. Likewise checking to be sure that the return value from the countMyFingers is an integer, and even happens at all (outside of a throw) is unnecessary because the compiler forces it.

Not the case for JavaScript, where both the data types on input and output, as well as whether or not the method returns a value could vary. In JavaScript, checking these conditions is an interesting test.

Use unit tests to protect behavior in situations where the language or platform do not offer protection.

Fragile data types

Datatypes that can build enforcement into them can offer protection and lift a burden on code that processes those datatypes. Other datatypes have no such capability, pushing the burden on code that processes them.

In the example I offered, the data type has no opportunity to provide such enforcement, as the data structure is built dynamically in the code before calling it. The relationship between category and data could be anything. It is very similar to arguments and inputs presented as strings, for example Json or XML, for that matter.

  var operation = { category: "execute_check", data: checkEntity};

Consider the following alternative:

class OperationInstructions
{
  public void OperationInstructions( CategoryEnum category, BaseCheckObject check)
  {
    if(category.Equals(CategoryEnum.ExecuteCheck && check.GetType().Equals(CheckEntity))
    {
      throw InvalidOperationException();
    }
  }
}

In this case, the input data cannot be created in the invalid state because it throws in the constructor. We can move the tests which protect against this condition to check the data type constructor and lift the burden on processOperation() for having to check at all.

We only have this option when the technology and platform we use permits it. For example, if implementing a REST API, the inputs are going to be in text format, and if the language we are using does not allow us to let the platform handle the data binding to strongly typed objects (as in C# and ASP.NET, for example) then we cannot exploit such mechanisms so easily.

When the data type offers no way to self-protect against invalid states, then use unit tests agains the code that processes the data to check for handling of invalid states.

and…

Consider changing the implementation to allow for data types that can either inherently offer checking and protection, or that change the problem in such a way that the opportunity for invalid data doesn’t exist.

Distant and far away protection

diagram of a three tiered system with client, api, and backend where data correctness is checked in the client

The system shown above is a nightmare scenario where we rely on the client to enforce correct data. Used for illustration only, not at all advised or prescribed.

Where is the guarantee of impossible state happening? Does the immediate caller of the method provide that guarantee? Is the caller of the method the only one that will use this method, or is the method available for other parts of the code inside and outside the component or even system, possibly going to call the method? Does that burden move likewise to all those callers?

Does the guarantee of impossible state happen further away, in another component or system? Is this a layered and distributed system? Is this protection happening in the backend, in the database, in the front end API layer, in web UI, or possibly even across the cloud and in a client or external system?

Which component is guaranteeing you never have the “impossible” state? The further away it is, the more possible other components there are, and the less control you have over that component, the greater risk you have that “impossible” isn’t true.

When the guard against the errant condition is far away from the code under test, and when multiple components may be doing the guarding, add unit tests to the code to likewise check the negative data condition.

Things change, and you don’t always understand the system. When we say something is impossible, what we really mean is “The system as I understand it, in the state it is in right now, cannot produce that state.” We can be wrong about the system, even if we built it. We also cannot know the future and how it may be different than the present, even if we are the ones that might change the future.

The more complex the behavior of the system, the requirements, the number of possible states, the more difficult it is for us to understand it. The more we plan to change something, the more of a future a given thing has and the more its use is growing, the more likely that some future change will make the currently impossible suddenly very probable.

If the system is complex, and if you anticipate a lot of change in the system, check handling of errant and negative inputs with a unit test.

The unit may be the only opportunity to test the impossible

Very often code handles enforcement of valid inputs in stacks. A unit of code deep in the stack may check for maximum and minimum range on values, or throw if given an invalid combination of inputs. The user interface, for sake of a better user experience, may check those inputs and inform the user of valid ranges, or might somehow disallow creation of those conditions.

diagram of a system with guards against invalid data in client and in backend, and showing how testing is blocked if only from client

We may still want to check that the unit of code guards against those conditions. If the layers above the unit provide similar guards, then the only way to test that behavior will be at the unit test level. You might have separate tests, one at the unit and one in the UI, but the purpose of the tests are different. The unit test is to detect if the business logic is allowing invalid states. The UI test is to detect if the UI logic is protecting the user experience.

This is one of those situations where the developer role, or at least someone taking on the developer responsibility of checking at the unit level, is the only role that can do a given test. When enforcement is low level in this way, the checking should happen that low as well. If it is not checked that low where the behavior happens, the responsibility to check ripples up the stack where the testing is far more difficult, slower, and as illustrated above, might be impossible.

When separate layers guard against invalid or errant conditions, the only way to check unit layer guards against invalid conditions is with negative tests at the unit level.

Possibly wrapping up now

When and how and what we ought to test for is not always an easy question to answer. It is even more difficult when we consider negative tests against conditions that we believe the system will never encounter. Whether or not such conditions are impossible is difficult to know, so you should take several factors into consideration when deciding how to handle such conditions and whether to write the unit test.

The factors that motivate considering a unit test are based on the risk of the “impossible” suddenly becoming possible, and the likelihood that other code will rely on the behavior of the code in the invalid data state. Through careful consideration we can build a set of checks that help us detect regressions before we submit the code, and through consideration of the testing we might even make design decisions that make the problem not even a consideration.