Zero is Random

Some observations on developing "random" implementations with Test Driven Development.

Random is "difficult to test" so at a minimum you want at least a seam between random and "complicated code that does something useful with random".  You really want all of the weird branching that you are responsible for maintaining to be hidden (in the Parnas 1971 sense) behind some facade that looks like a bits to useful result converter with deterministic behavior.

The one input that you know for certain you'll need to handle gracefully occurs when all of the random bits are false, so unless you have a compelling reason to do something else, your first test should be "what value do we get back if we pass in zero?"

You also need to pick an expected value to be returned.  You've got some degrees of freedom available, so you should spend some of them by making the zero case that result with is most convenient to assign to zero.

I normally reach for squashed encoding as my mapping of choice, so for me the zero case is taking the first available option at every decision point.  For the Fisher chess problem, each of the bishops is assigned to the first available candidate square.  For room asignments in Hunt the Wumpus, the hunter is placed in room 1, the wumpus in room 2, and so on.

In my practice, a single input variable can surface enough duplication that you can start refactoring right away (when I was first learning TDD, I didn't realize this, but Kent pointed out the duplication and the lesson stuck).

But if your practice of TDD doesn't allow you to change your code without a broken test, then the next candidate input value to consider is 1, because you will always need to handle gracefully the case where the least significant bit is true, and everything else is false.  And your result of choice will be almost exactly the same as the zero result, with the first branch taken differently.

For Hunt the Wumpus, this might mean switching the room assignments of the hunter and the wumpus, while leaving the rest of the room assignments unchanged.

TDDbE: Mastering TDD

In TDD, the tests are a means to an end -- the end being code in which we have great confidence.  If our knowledge of the implementation gives us confidence even without a test, then we will not write that test.

This is, I think, an important piece of the design effort; we need to recognize where the parts of the system are where testing will have unacceptable ROI, and build into our designs the constraint that these parts of the system must be expressed using code that is "so simple there are obviously no deficiencies".

It's the complicated parts of the code (Conditionals, Loops, Operations, Polymorphism) where we need the testability constraint, and a test first approach is an acceptable mechanism for fixing that constraint.

If we see parts of the system that could be dramatically simplified but that don't demand change at the moment, then we will leave them alone.

Adam Tornhill's work suggests that we should be using the source control system to identify hotspots.  If TDD is going to make a difference, it has to be at the points where the infelicities of the previous design have real impact on daily work.

TDD helps you to pay attention to the right issues at the right time so you can make your designs cleaner, you can refine your designs as you learn....  What's naive is assuming that clean code is all there is to success.

The Chrysler project was cancelled; as far as I can tell LifeWare is still around.

Mike Hill points out that for the past 25 or whatever years, the software development business has been over provisioned in money, compute, and resources.  If that gets squeezed back down to "normal" levels, then design practices may become much more significant to long term success.

 

TDDbE: Refactoring

In TDD we use refactoring in an interesting way.  Usually, a refactoring cannot change the semantics of the program under any circumstances.  In TDD, the circumstances we care about are the tests that are already passing.  So, for example, we can replace constants with variables in TDD and, in good conscience, call this operation a refactoring, because it doesn't change the set of tests that pass.

Once upon a time, there was a video interview with Martin Fowler where he was vehement that changing any observable behavior -- even those which you weren't currently observing -- was out of bounds.

For example, when starting new work, we often complete the GREEN task by returning a fixed constant.  "Everybody" agrees that you can now clean up the implementation as much as you like, so long as the function returns that fixed constant regardless of your inputs.

But this often means data duplication - the constant hard coded into the implementation is a copy of information that lives in the test; the expected return value of the function.

If you think removing this duplication is important, then maybe you aren't going to care so much about the refactoring rule that says that all observable behaviors must be preserved.

On the other hand, if you think the refactoring rule is important, well... it's not like implementing a second test is all that far out of your way, and if you are clever about the refactoring you do beforehand, you won't need very much wall clock time at all to get this second test passing.

My view is that I don't need permission from the test to improve the code ("the code works for me, I don't work for the code").  And if a purist I'm pairing with object that such a change isn't a refactoring - fine, I won't call it a refactoring..., but I'm going to make a case for doing it whatever we call it.

Leap-of-faith refactoring is exactly what we're trying to avoid with our strategy of small steps and concrete feedback.

Mikado method might be a useful technique here - try something that could work, in the event of negative feedback revert the change, and start building out the graph of small changes that you need to make to get to the end.

First, isolate the part that has to change.

Which calls back to Parnas on Information Hiding, or the advice to write code that is easy to delete.

How do you simplify control flows that have become too twisted or scattered?  Replace a method invocation with the method itself.

This is a Big Power Move.  Even if you ultimately revert the code back to its original state, being able to see what the program is doing, released from the abstractions obscuring that, can be a huge time saver.  See also Carmack 1997.

TDDbE: Design Patterns

This chapter gives us a short catalog of design patterns.  Beck offers an interesting insight, which he credits to Alexander: patterns are a consequence of the fact that we keep using the same tools to solve different problems.

Design in TDD requires a slightly different look at design patterns.

For those keeping score, that's now three things where Kent's concepts have a slightly different take.

He separates his catalog into patterns that appear in test writing, patterns that appear during refactoring, and patterns that can appear in both contexts.  Okay... but it's not at all obvious to me that his clusters reflect where these patterns appear in my execution of TDD (why wouldn't value objects be used during refactoring)?

I find the descriptions of the patterns to be little more than sketches; in particular, I find it difficult to understand how some of these patterns differ from the earlier treatments, beyond the fact that he uses Java, rather than Smalltalk or C++ as his primary examples.

Finding Imposters during refactoring is driven by eliminating duplication, just as all refactoring is driven by eliminating duplication.

This I have to dismiss as overstatement, of the same nature as "always start with an automated test".  Changing the code to look like we knew what we were doing all along is a perfectly reasonable motivation for refactoring, even if our current design has already eliminated all of the "duplication".

 

 

TDDbE: xUnit Patterns

 Much more so than the previous chapters in this book, this one feels like archeology.

Write boolean expressions that automate your judgment about whether the code worked.

Primarily, the point of the automated mistake detectors is that they don't require sapience to reach a conclusion (see Back and Bolton on checking vs testing).  There shouldn't be any maybe or fuzzing of the outcomes; we're on the happy path, or we're not.

The expected value generally goes first....

Today, I read this as an indicator that the right abstractions hadn't yet emerged; we were all still thinking about the "two arguments to assert", rather than say a measurement that needs to satisfy a specification.  And of course some frameworks flipped the arguments, and finding that programmers have chosen an order that is the reverse of what the current framework uses is a common error.

I'm aware that I am swimming against the tide in insisting that all tests be written using on public protocol.

I think is "all" is a bit too extreme; also, this doesn't surprise me.  You can go back and forth on the various trade offs, but ultimately it comes down again to the fact that the tests work for the developers, not the other way around.

I was interested to note the remarks about dropping into the Smalltalk debugger, as this same behavior was an important motivation for Mock Objects.

Each new kind of fixture should be a new subclass of TestCase.  Each new fixture is created in an instance of that subclass, used once, and then discarded.

I haven't used "fixtures", especially fixtures that require explicit setup/teardown, in a long long time.

xUnit circa 2003 implemented fixtures using overridden class methods because... well, because most of the modern alternatives hadn't been realized yet.  I don't love that design, because I don't think it is obvious enough that the fixture is being re-initialized for every test call, so the class design looks like shared potentially mutable state.

Test "outlines" are now more commonly realized via nesting of contexts, so that you get a report that is a hierarchy of test results.

My experiences suggest that TestSuites are extinct in languages where the test running has access to class metadata, because everything is discovered, often with an annotation language that allows for enabling and disabling tests, or mapping tests against a wider variety of inputs, or whatever.

In summary: languages, test frameworks, and development environments have made a lot of real progress over the past 20 years; the details of how we made do without modern conveniences aren't particularly compelling.

TDDbE: Green Bar Patterns

The chapter opens with a discussion of Fake It, Triangulate, and a light comparison of the two.

Does Fake It violate the rule that says you don't write any code that isn't needed?  I don't think so, because in the refactoring step you are eliminating duplication of data between the test case and the code.

More support for my favorite lesson: duplication of data between the test case and the code is all the permission you need to continue refactoring.

I only use Triangulation when I'm really really unsure about the correct abstraction for the calculation.

I think Triangulation really suffers from the fact that the examples typically used to demonstrate it are contrived.

Beck uses addition here, he used an equality check back in chapter three -- both are cases where you could just type in the correct implementation, rather than faffing about with triangulating.

The sorts of problems that I would expect to benefit from triangulating would be more like sudoku, or code formatting, or even line wrapping, where you may need to work in the code for a bit before you get the key insight that gives you confidence in a specific abstraction.

Sandi Metz demonstrates with "simple green" that you might prefer the duplication to the abstraction anyway.  Her advice was to prefer duplication over the wrong abstraction.

Back in Chapter 17 Kent observed that the money example went in "a completely different direction".  And as I understand it he didn't "triangulate" to choose between candidate designs - he discovered a metaphor and just went where it lead him.  This tells me that triangulation is more about the uncertainty that the abstraction will in fact be better than the duplication.

Of course, you could also try the abstraction, and then use the source control system to revert to a checkpoint if you decide that the refactoring hasn't led to an abstraction you trust.

There's no particular virtue in the halfway nature of Fake It and Triangulation.

There ceremonies are a means to an end....

  Keep track of how often you are surprised by red bars....

"Call your shots" later became a more general thing - the earliest reference I can find is Needham 2010, but he in turn attributes the idea to Kent.

TDDbE: Testing Patterns

When I write a test that is too big, I first try to learn the lesson.  Why was it too big? What could I have done differently that would have made it smaller? ... I delete the offending test and start over.

Delete and start over is a power move, and will later lead to the exploration of Test Commit/Revert.

This reminds me of Mikado method: that we might want to systematically reset and explore smaller goals until we get to an improvement we actually can make now, and then work our way back up the graph until the code is ready for the big test.

I think you have to be fairly far along in the design to start running into this sort of thing; when the design is shallow, you can usually just drop in a guard clause and early return, which gets the test passing, and now it is "just" a matter of integrating this new logic into the rest of the design.  But when the current output is the result of information crossing many barriers, getting the guard clause to work can require rework across several boundaries.

How do you test an object that relies on an expensive of complicated resource?

Oh dear.

OK, context: Beck is writing prior to the development of the pattern language of Test Doubles by Gerard Meszaros.  He's re-purposing the label introduced in the Endo-Testing paper, but it's not actually all that great a fit.  It would be better to talk about test stubs (Binder, 1999), or doubles if you wanted to be understood by a modern audience.

Anyway, we get a couple different examples of stubs/doubles that you might find useful.

Leaving a test RED to resume at the next session is a pretty good trick; but I think you'll do better to treat the test as a suggestion, rather than a commitment, to resume work at a particular place - the most important item to work next might change between sessions, and the failing test will be a minor obstacle if you decide that refactoring is worthwhile.