It’s awfully presumptuous for me to be writing this. After all, I only really started test-driving my code a year ago, and even then it was because I needed to demonstrate fanatical, fundamentalist, write-no-code-except-to-turn-the-red-bar-green Kent Beck TDD to get a job. I threw myself into the deep end, got the job (and it’s a great job!), and somewhere along the line realized that red-green-refactor had incorporated itself into my idea of what constitutes “good enough”. So this is to a not insignificant extent the zeal of the convert; let’s all take it with a pillar of salt.

Or maybe let’s not. I’ve worked with a lot of people who claim to test-drive their code but mostly just let code reviews drive their tests; for them, unit tests are something you write because your pull request won’t pass review without them, a chore you complete to the minimum possible bar of effort. And you’ve got fifteen endorsements for “TDD” on your LinkedIn profile, because of course you do.

I’ve worked with even more people who “just don’t see what all the fuss is about”. Testing is QA’s job, isn’t it? They have massive suites of regression tests that take days to run, that’s hardly a practical way to write code. And what’s the point in writing a test if it won’t even compile?

These people often tell me that TDD is an “academic” exercise. I am here to tell you that I crashed and burned with TDD when I was an academic, probably because I was doing it wrong and probably because the tooling was way behind the level I currently enjoy. Mainly it was because my whole department held it in contempt – what sort of bold academic needs a safety net to climb onto the shoulders of giants? – and it’s difficult to pursue something, badly, when the people who’re supposed to be your role models are telling you it’s a foolish and cowardly waste of time.

Anyway, let’s get on with the programme and discover if you really don’t need to test-drive your code.

TDD to get rapid feedback

Imagine writing code for hours, churning out dozens of methods and hundreds of lines, until you finally get to a place where your code will build and run – and then discovering that it’s all wrong. You diligently step through it in the debugger; the NullReferenceExceptions are easy to find and fix (except when there’s no way that reference could be null, not here!), but you find yourself groveling over the values of locals as they flicker from iteration to loop iteration, trying to sort signal from noise.

What a nightmare! Good thing it only took me eighteen years of programming experience to stop doing that on a professional basis.

A decent unit testing framework lets you exercise even the innermost reaches of your code as soon as you’ve written it. You don’t need an entry point, or a user interface, or a SOA framework to drive the code that calls into your domain logic – you just poke the domain code with a properly-shaped stick and see if it gives you the right kind of oink. “Hey, does my code actually work?” Find out in five seconds.

Remco Mulder’s NCrunch is a minor miracle in this regard. Remembering to hit ctrl-U, U after fixing up a few lines of business logic, then waiting for the project to build and your tests to run, doesn’t seem like a hardship until you’ve had a piece of code take care of it for you. If you’re not a .NET partisan, I’m sure you can still find continuous-testing tools for your platform of choice. You should use them.

Maybe don’t TDD if you live in a REPL

TDD’s biggest advantage, as far as I’m concerned, is that it enables damn near instant feedback about the state of your code. If you constantly run and exercise your code in a REPL, maybe you’re already getting instant feedback!

Of course, coding in a REPL and test-driving your code aren’t mutually exclusive. Maybe you nail down boundary conditions or acceptance criteria with automated tests so you don’t have to set them up by hand every time you change your code, but throw ad-hoc test data at your code in the REPL when it’s easier to do that than to code up the scaffold of yet another damn test case. Like beer and brisket, REPLs and TDD are two great tastes that taste great together.

Definitely don’t TDD if your build times are horrifying

I know how it is, I’ve worked on four gigabytes of C++ codebase too. If a best-case incremental build takes you ten minutes and touching a header file means come back tomorrow, you probably won’t get much out of fundamentalist TDD. Maybe you can use a test framework to execute some of your code before the whole ponderous beast is assembled, but at the very best your step sizes are likely to be a lot longer.

If you’ve done tight-loop TDD on a project with full build times measured in hours, by all means yell at me on twitter. Better yet, blog about it and show me up by posting it for all to see!

TDD to document your code in a way that cannot lie

The problem with code-level documentation – or any documentation, for that matter – is that it’s a pain in the ass to write and rapidly becomes irrelevant as the code evolves. I’ve written a ton of wiki doc, mostly because that was the least bad option. It’s hard to write, hard to find, hard to relate to the code, and requires constant upkeep. Even then, maybe thirty percent of the people who ought to read it ever will.

Tests work much better as documentation. If you’re writing code, you should already have the tests available to you – they’re in the repo you checked out, probably in the same project as the code you’re working on. They’re easy to find (anything from “Find all references” to grep will do the job), and if they pass they’ll tell you exactly how the code in question behaves.

If you write your tests first, you’re blessed with not having to remember all the features you have to document, because you documented them by test-driving them. You also won’t have to remember to document all the corner cases, because you handled those by test-driving them, too.

Don’t rely entirely on TDD if your docs require bling

If you need flow-charts or UML or Powerpoint to hit your requirements, TDD isn’t your silver bullet.

Don’t count on TDD for documenting an open host

If you want to document your bounded context’s interface with the rest of the big, wide, scary, uncaring world… unit tests aren’t gonna cut it. Probably you’ve come across something like Swagger or Apiary if you have an HTTP REST API, and they aren’t notably terrible. Perhaps you’ll be able to use Pacts or something similar. That sounds nice. Can I come visit?

Definitely don’t TDD if you’re a Real Programmer

“If it was hard to write, it should be hard to understand!”

Also, tests as documentation will deprive you of the hilarious fun of comments like:

i--; // Add one to i

Also also, I hope we never meet. Jerk.

TDD to write more SOLID code

For all that I’m an ardent advocate of test-driven development, I’m still a bit skeptical about test-driven design. Maybe I’ll come around at some point. But either way, test-driving your code makes it more natural to write code that’s well-separated, just because it’s easier to test chunks of code that only have to do one thing. It makes it more natural to write loosely-coupled code, because if two classes are close collaborators they get painful to unit test in isolation. You might find yourself tending towards better command-query separation, because queries are easy to test and commands are bad enough without having to check for output as well.

This might sound familiar!

Don’t TDD if you don’t know when to stop

David Heinemeier-Hansson coined the term Test-Induced Design Damage. We should rightly fear TIDD Syndrome, where every collaborator can be mocked out by an interface (you need an interface, because you’ll need to mock it for tests!) and every operation is performed by instantiating a class that satisfies a particular interface that provides a particular service. I’m exaggerating, but holy shit, not by much.

One ominous sign of TIDD is when you ask yourself, or your colleague, or Stack Overflow: “How do I test my DI container configuration?” At that point you should probably just set your codebase on fire before it infects the whole city.

If you find yourself with many interlocking patterns of mutable and interdependent state – in DDD terms, that means a thundering herd of entities that can’t not be in the same aggregate – you might have a problem. If you find that you have to inject seven or eight different mock objects into a single class in order to test one of its public methods, that class might be doing too much. If you keep adding stuff to a single kitchen-sink interface because you don’t want to mock up yet another goddamn object, you might consider taking a deep breath and a step back. All the pain that “test-driven development” is causing you just might be a suggestion to rethink your design.

I’d offer you a solution, but I don’t have to, because Mark Seeman has already done a damn good job on PluralSight. It turns out that, rather than depending upon a complex web of invariants between various chunks of mutable state, it’s a lot easier if you can make the mutable-state part go away.

And finally, we get to:

TDD to do your fucking job

Yes, obviously it’s possible to do your job without test-driving your code. I’ve done it for well over a decade, and it was more trouble than it had to be (although I was mostly writing C and C++, so maybe I’m wrong about that – see “compile times”, above). But if you give a shit, if you take some pride in what you do, you don’t like getting test-case fails back from QA or having some smug asshole (or dear colleague) point out obvious bugs in code review. You want to push your code and raise a pull request in unshakeable certainty that you’re going to bathe in the admiration of everyone who reads it, and that when your ticket hits QA it’ll sail through in minutes.

Since I started religiously (yeah, the term is indicative, sorry I’m not sorry) test-driving my code, I tend to forget to actually test it by hand. I don’t really bother firing up Postman and verifying the ticket’s acceptance criteria.

Maybe one time in fifteen I miss a spot and I get rapped across the knuckles by QA for my negligence. That just means I need to write better tests. With the panoply of test-automation frameworks we have available, it’s ridiculous that anyone should have to fiddle-fuck HTTP request parameters or hand-hack JSON payloads to check my work against the acceptance criteria that we’d originally written into the ticket. That’s busy-work at best.

Don’t TDD if you enjoy goldbricking

“Oh, I thought I had all the acceptance criteria covered, I guess I missed that one.” Shit, now you’ll have to fix that one thing and put it back into code review and wait until QA has time to pick it up again. That ought to give you plenty of time for Facebook.

Also, I hope we never meet. Jerk.

Definitely TDD if you enjoy goldbricking a little too much

If you give me a clear board with no new tasks to pick up, I’ll find myself something to do. There’s always code that needs cleaned up, or a utility library that should probably be documented so we can quit reinventing its functionality every time we write up a test suite for a new project. I’ll get there, eventually. Might take a minute to check Twitter first. Shit, am I out of coffee?

If you show me a test that’s red, I turn into a missile. I’ll cheerfully spend the next four hours tracking down that end-to-end test failure to that submodule update that got applied here but not over there, and I don’t care what just happened in free agency. I’ll catch up on the train home, nbd, I have work to do.

Red-green-refactor hits the same spot for me. I can’t let go of a failing test, and once I get it to pass I can’t usually bring myself to push that commit until I’ve refactored it to not suck. I expect you’re the same way, because this is (broadly speaking) how slot machines and MMOs work.

Appendix: TDD your changes to legacy code

“Get code under test before refactoring” – I really don’t have to explain this, do I? Granted, this is more a matter of test-_first_ development than test-_driven_ development, because you’re starting out by nailing down the behaviour of the big ball of mud you need to modify rather than writing failing tests to describe its new behaviour, but it’s far better than blindly messing with existing code and hoping it all works out for the best.

Don’t TDD your changes if it’s easier to regression-test

Sometimes it’s straightforward to run an automated regression test suite against your codebase, but breaking enough dependencies to get the particular module you’re hacking on properly under test would take weeks. If you don’t have weeks (or, maybe, if you do have weeks but nobody’s likely to benefit from the fruits of your labour), labouring to inject a zillion dependency interfaces just so you can abuse your favourite mocking library is not the most productive use of your time.

Fiddle-fucking around with partial mocks and argument matching and so on can be a lot of fun, in that peculiar complexity-worshipping way we tend to enjoy, but let’s not underestimate the power of throwing a brute-force regression suite at a piece of legacy code on a modern, high-horsepower computer. Sometimes it takes less clock time than all the reflection your mocking framework needs to get itself bootstrapped.