Srikanth Sastry A Techie in Boston

The merits of unit tests — Part 3

Cross posted on LinkedIn.

Previously, I talked about how unit tests serve purposes other than verifying code functionality; I talked about unit tests serving as defacto documentation, and unit tests helping you refactor without fear. In this post, I'll talk about yet another benefit to unit tests: writing better software.

Better software design

I am not talking about writing bug-free software here. Sure, good unit tests help you discover/avoid large classes of bugs, but that's not the point. Unit tests also help you enforce good design patterns and modularity in your software.

Unit tests help you with software design in two complementary ways. First, they help you establish optimal boundaries of modularity in terms of methods, and classes, and second, they help you understand your dependencies better and almost force you to use good dependency injection hygiene.

Modularity

How do you know that your class or method does 'too much', or that it has undesirable side-effects? A pretty good way to discover it is to start writing unit tests for it. If you find yourself having to test for too many different types of inputs, then your methods are doing too much. If you find yourself having to test for too many orderings of operations, when your methods have too many side effects. It really is just as simple as that!

For instance, some time ago, I came across a piece of code that essentially took a large query, split it up into multiple subqueries, sent them off in parallel, collected their responses back, munged them and returned the munged response to the caller via a callback. The code looked something like this:

public class QueryManager {
void sendRequest(Query query, Callback queryCallback) {
List<Subquery> subqueries = splitQuery(query);
for (Subquery subquery : subqueries) {
sendQuery(subquery, new Subquery.Callback() {
@Override
public onSuccess(Response response) {
// Do some processing. ...
if (allResponses()) {
queryCallback.finalResult();
} else {
queryCallback.incrementalResult();
}
}
@Override
public onFailure(Error e) {
// Do some error handling. ...
if (allResponses()) {
queryCallback.finalResult();
}
}
});
}
}
// Other methods. ...
}

This code makes for an interesting case study on multiple fronts, and I will come back to it in a later post. For now, it is sufficient to state that I wanted to make changes to this code, but it was pretty thorny because (you guessed it!) it has no unit tests.

Naturally, the first step is to write unit tests for this class, and that was when I realized why there were no unit tests here. This class is incredibly tricky to unit test. Getting the subqueries to respond under various conditions and ordering resulted in a combinatorial explosion of test cases making the task intractable. This was very strong code smell.

As you have probably figured out already, this was a classic case of a single class doing too much. The primary culprit was the anonymous class that implemented the Subquery.Callback interface. It should really have been its own class with its own unit tests.

After I pulled out that anonymous class into its own class, it became a lot easier to unit test both the Subquery.Callback and the QueryManager individually, and with that, the code became much more modular, easier to read, and much easier to maintain.

Dependency

If your code does not do a decent job of injecting its dependencies from outside, you are gonna have a bad time! Having good unit tests will actually keep you from getting into this pitfall pretty effectively. Consider the following contrived example. You have a piece of code that writes to an external service, and your code throttles the rate of writes because going over your approved rate/quota can be pretty expensive. So, your code could look something like this:

class RateLimiter {
void writeToExternalService(const vector<Entries>& stuff) {
auto service = new ExternalService(ConnectionParameters foo);
for (auto entry : stuff) {
waitUntilQuotaAvailable();
service.write(entry);
}
}
}

Remember how I said that going over the rate/quota is bad? How do you verify that it will not happen? Well, you could set up an elaborate testbed that has an ExternalService simulator, and you can run your code through all sorts of inputs and verify that the simulator says that the rate limiting is work. But that's expensive, and if you choose to go with a different external service, then well, good luck with that!

Instead, you could try to unit test it. But how? You need to have access to the ExternalService to do that, which we have already established is expensive! Well, this is where unit testing it will force you into healthy dependency injection. For this contrived example, you can inject the dependency as follows.

class RateLimiter {
void writeToExternalService(
const vector<Entries>& stuff,
ExternalService service) {
for (auto entry : stuff) {
waitUntilQuotaAvailable();
service.write(entry);
}
}
}

Your unit tests could do this:

TEST(AwesomeTestCase, testWrite) {
vector<Entries> stuff = generateLotsOfStuff();
RateLimits limit; // Set your limits here.
ExternalService svc = MockExternalService(limit);
RateLimiter rateLimiter = new RateLimiter(limit);
rateLimiter.writeToExternalService(stuff, svc);
ASSERT_TRUE(svc.writesWithinLimits());
}

And you can put your implementation through the wringer without breaking a sweat.