This book is about writing cost-effective, maintainable, and pleasing code.

Chapter 1 explores how to decide if code is "good enough." This chapter uses metrics to compare several possible solutions to the 99 Bottles problem. It introduces a type of solution known as Shameless Green, and argues that although Shameless Green is neither clever nor changeable, it is the best initial solution to many problems.

Chapter 2 is a primer for Test-Driven Development (TDD), which is used to find Shameless Green. This chapter is concerned with deciding what to test, and with creating tests that happily tolerate changes to the underlying code.

Chapter 3 introduces a new requirement (six-pack), which leads to a discussion of how to decide where to start when changing code. This chapter examines the Open/Closed Principle, and then explores code smells. The chapter then defines a simple set of Flocking Rules, which guide a step-by-step refactoring of code.

Chapter 4 continues the step-by-step refactoring begun in Chapter 3. It iteratively applies the Flocking Rules, eventually stumbles across the need for the Liskov Substitution Principle, and ultimately unearths a deeply hidden abstraction.

Chapter 5 inventories the existing code for smells, chooses the most prominent one, and uses it to trigger the creation of a new class. Along the way, it takes a hard look at immutability, performance, and caching.

Chapter 6 performs a miracle that not only removes the conditionals, but also allows you to finally implement the new six-pack requirement without altering existing code.

Chapter 7 examines the tradeoffs along a continuum of six different styles of Factories. It begins by exploring a simple, hard-coded conditional, and ends with a factory whose candidate objects both self-register, and also supply the logic needed to choose them.

Chapter 8 introduces another new requirement—​to vary the lyrics. It uses this requirement to introduce the idea of a programming aesthetic, or set of rules to guide you in times of uncertainty. The chapter ends with a list of specific suggestions for deciding when it’s worthwhile to voluntarily improve code.

Chapter 9 comes full circle and returns to testing. It takes advantage of the improved design to write better tests, and then uses the new tests as a spur to improve the final design.

The lessons in the book have been found useful by programmers with a broad range of experience, from rank novice through grizzled veteran. Despite what one might predict, novices often have an easier time with this material. As they are unencumbered by prior knowledge, their minds are open, and easily absorb these ideas.

It’s the veterans who struggle. Their habits are deeply ingrained. They know themselves to be good at programming. They feel quick, and efficient, and so resist new techniques, especially when those techniques temporarily slow them down.

This book will be useful if you are a veteran, but it cannot be denied that it teaches programming techniques that likely contradict your current practice. Changing entrenched ideas can be painful. However, you cannot make informed decisions about the value of new ideas unless you thoroughly understand them, and to understand them you must commit, wholeheartedly, to learning them.

Therefore, if you are a veteran, it’s best to adopt the novice mindset before reading on. Set aside prior beliefs, and dedicate yourself to what follows. While reading, resist the urge to resist. Read the entire book, work the problems, and only then decide whether to integrate these ideas into your daily practice.

You’ll learn more from this book if you spend 30 minutes working on the "99 Bottles of Beer" problem before starting to read. See the appendix for instructions.

If you just want to read on but you don’t know PHP, have no fear. Your purchase of this book entitles you to all variants (Ruby, JavaScript, PHP, and Python), and each language variant comes in beer and milk flavors. Every combination of variant and flavor was available for download during your purchase, and one of them should suit.

Regardless of which version you read, be assured that this book is not about any specific language or beverage; it’s about object-oriented programming and design. The technical content of every version is essentially the same.

The chapters build upon one another, and so should be read in order. While isolated sections may be useful, the whole is more than the sum of its parts. The ideas gain power in relation to one another.

To get the most from the book, work the code samples as you read along. With active participation, you’ll learn more, understand better, and retain longer. While reading has value, doing has more.

The exercises rely on PHPUnit.

A current list of errata is located at sandimetz.com/99bottles-errata. If you find additional errors, please email them to errata@99bottlesbook.com.

Sandi Metz

Sandi is the author of Practical Object-Oriented Design in Ruby. She has thirty years of experience working on large object-oriented applications. She’s spoken about programming, object-oriented design and refactoring at numerous conferences including Agile Alliance Technical Conference, Craft Conf, Øredev, RailsConf, and RubyConf. She believes in simple code and straightforward explanations, and is the proud recipient of a Ruby Hero award for her contribution to the Ruby community. She prefers working software, practical solutions and lengthy bicycle trips (not necessarily in that order). Find out more about Sandi at sandimetz.com.

Katrina Owen

Katrina works for GitHub as an Advocate on the Open Source team. Katrina has ten years of experience and works primarily in Go and Ruby. She is the creator of exercism.org, a platform for programming skill development in more than 30 languages. She’s spoken about refactoring and open source at international conferences such as NordicRuby, Mix-IT, Software Craftsmanship North America, OSCON, Bath Ruby and RailsConf. She received a Ruby Hero award for her contribution to the Ruby community. When programming, her focus is on automation, workflow optimization, and refactoring. Find out more about Katrina at exercism.github.io/kytrinyx.

TJ Stankus

TJ works as a software developer for boldfacet LLC and co-instructs software design courses with Sandi. He began his programming career over 20 years ago by accident. By hacking together WordPerfect macros to streamline his job as a proofreader, he discovered he loved programming as much as any creative activity he’d ever pursued. He has worked in mobile applications and back-end web development. He even wrote an SMTP server back when that seemed like a good idea. Today, he works primarily with Elixir and Ruby. His main interests lie not in specific programming languages, but in the essential design ideas that span programming languages and paradigms. Find out more about TJ at tj.stank.us.

Tom Stuart

The JavaScript translation was done by Tom Stuart. Tom is a computer scientist, programmer and technical leader. He’s the former CTO of FutureLearn and Econsultancy. He has lectured on optimising compilers at the University of Cambridge and written about technology for the Guardian.

Tom is the proud author of Understanding Computation, which was published by O’Reilly in 2013.

Matthew Fonda

Matthew Fonda provided deeply thoughtful and extremely helpful guidance about how to think about PHP and OO. Matthew is CTO at eNotes.com, where he has spent the past 15 years building a platform to help students learn. During his time at eNotes, he has learned firsthand the importance of building codebases that are easy to change over time, and is more certain than ever that the answer to every question in software development is "it depends". Matthew is a longtime member of the PHP community and occasional contributor to PHP’s documentation. When not at the computer, you can find Matthew snowboarding in the mountains of the Pacific Northwest.

The Stocktons

Brothers Dave and Dann Stockton spent countless hours pondering the initial PHP translation of the 1st edition. Theirs was the first attempt at transforming the book to use another programming language, and as such they suffered all the pain one might imagine. We are eternally grateful for their efforts and contributions.

The PHP shown in the book was written by Katrina, with the generous input of Matthew, Dave and Dann.

Katrina Owen Katrina Owen is not only an author of this book, but also wrote the Python translation.

The code you write should meet two often-contradictory goals. It must remain concrete enough to be understood while simultaneously being abstract enough to allow for change.

Imagine a continuum with "most concrete" at one end and "most abstract" at the other. Code at the concrete end might be expressed as a single long procedure full of if statements. Code at the abstract end might consist of many classes, each with one method containing a single line of code.

The best solution for most problems lies not at the extremes of this continuum, but somewhere in the middle. There’s a sweet spot that represents the perfect compromise between comprehension and changeability, and it’s your job as a programmer to find it.

This section discusses four different solutions to the "99 Bottles of Beer" problem. These solutions vary in complexity and thus illustrate different points along this continuum.

You must now make a decision. As you were forewarned in the preface, the best way to learn from this book is to work the exercises yourself. If you continue reading before solving the problem in your own way, your ideas will be contaminated by the code that follows. Therefore, if you plan to work along, go do the 99 Bottles exercise now. When you’re finished, you’ll be ready to examine the following four solutions.

You now have access to five different solutions to the "99 Bottles of Beer" problem; the four listed in the preceding section and the one you wrote yourself.

Which is best?

You likely have an opinion on this question—one which, granted, may have been swayed by the commentary above. However, independent of that gentle influence, the sum of your experiences and expectations predispose you to assess the goodness of code in your own unique way.

You judge code constantly. Writing code requires making choices; the choices you make reflect personal, internalized criteria. You intend to write "good" code and if, in your estimation, you’ve written "bad" code, you are clearly aware that you’ve done so. Regardless of how implicit, unachievable, or unhelpful they may be, you already have rules about code.

While having standards of any sort is a virtue, the chance of achieving your standards is improved if they are explicit and quantifiable. Answering the question "What makes code good?" thus requires defining goodness in concrete and actionable ways.

This is harder than one might think.

As programmers grow, they get better at solving challenging problems, and become comfortable with complexity. This higher level of comfort sometimes leads to the belief that complexity is inevitable, as if it’s the natural, inescapable state of all finished code. However, there’s something beyond complexity—a higher level of simplicity. Infinitely experienced programmers do not write infinitely complex code; they write code that’s blindingly simple.

This chapter examined four possible solutions to the "99 Bottles" problem as a prelude to defining what it means to write simple code. It used metrics as a starting point, injected a bit of common sense, and landed on Shameless Green.

Shameless Green is defined as the solution that quickly reaches green while prioritizing understandability over changeability. It uses tests to drive comprehension, and patiently accumulates concrete examples while awaiting insight into underlying abstractions. It doesn’t dispute that DRY is good, rather, it believes that it is cheaper to manage temporary duplication than to recover from incorrect abstractions.

Writing Shameless Green is fast, and the resulting code might be "good enough." Most programmers find it embarrassingly duplicative, and the code is certainly not very object-oriented. However, if nothing ever changes, the most cost-effective strategy is to deploy this code and walk away.

The challenge comes when a change request arrives. Code that’s good enough when nothing ever changes may not be good enough when things do. Chapter 3 introduces just such a change, and in that chapter you’ll begin improving the code. Before moving on, however, it’s time to take a step back, and learn how to test-drive Shameless Green.

A generation ago, a handful of extreme programming (XP) practitioners began writing automated tests using a technique they called "test first development." Their ideas were so influential that automated tests are now the norm, and these tests are often written first, in prelude to writing code.

The practice of writing tests before writing code became known as test-driven development (TDD). In its simplest form, TDD works like this:

In Test-Driven Development by Example, Kent Beck describes this as the Red/Green/Refactor cycle and calls it "the TDD mantra."

The ideas of testing, and of testing first, have won the hearts and minds of programmers. However, a commitment to writing tests doesn’t make this easy. TDD presents a never-ending challenge. You must repeatedly decide which test to write next, how to arrange code so that the test passes, and how much refactoring to do once it does. Each decision requires judgment and has consequences.

If your TDD judgment is not yet fully developed, it’s reasonable to temporarily adopt that of a master. Here’s an excellent guiding principle:

Green means safety. Green indicates that, at least as evidenced by the tests at hand, you understand the problem. Green is the wall at your back that lets you move forward with confidence. Getting to green quickly simplifies all that follows.

This chapter illustrates how to incrementally create tests and then use these tests to drive the development of code. The examples obediently follow the Red/Green/Refactor cycle, but are fairly conservative. Because the initial goal is more about reaching green than writing perfect code, the refactoring step sometimes removes duplication and other times retains it.

The plan is to create tests that thoroughly describe the "99 Bottles" problem, and then to solve the problem with the implementation known as Shameless Green. The Shameless Green solution strives for maximum understandability but is generally unconcerned with changeability. Shameless Green does not assert that changeability isn’t important; it merely recognizes that getting to green quickly is often at odds with writing perfectly changeable code. This chapter concentrates on creating the tests and writing simple code to pass them. Future chapters refactor the resulting code to improve the design.

The first test is often the most difficult to write. At this point, you know the least about whatever it is you intend to do. Your problem is a big, fuzzy, amorphous blob, and it’s challenging to reach in and carve off a single piece. It feels important to choose well, because where you start informs how you’ll proceed, and ultimately determines where you’ll end. The first test can therefore seem fraught with peril.

Despite its apparent import, the choice you make here hardly matters. In the beginning, you have ideas about the problem but actually know very little. Your ideas may turn out to be correct, but it’s just as possible that time will prove them wrong. You can’t figure out what’s right until you write some tests, at which time you may realize that the best course of action is to throw everything away and start over. Therefore, the purpose of some of your tests might very well be to prove that they represent bad ideas. Learning which ideas won’t work is forward progress, however disappointing it may be in the moment.

So, while it is important to consider the problem and to sketch out an overall plan before writing the first test, don’t overthink it. Find a starting place and get going, in faith that as you proceed, the fog will clear.

If you were to sketch out a public Application Programming Interface (API) for "99 Bottles," it might look like this:

This API allows others to request a single verse, a range of verses, or the entire song.

Now that you have a plan for the API, there are a number of possibilities for the first test. You could write a test for the entire song, for a series of contiguous verses, or for any single verse. Because the easiest way to get started is to tackle something that you thoroughly understand, it makes sense to begin by testing a single verse, and the most logical first verse to test is the first verse to be sung. Here’s that test, written in PHPUnit:

This test, like all tests, contains three parts:

Lines 3-7 above define the expected result and are thus part of the setup. Setup continues on line 8, where a new bottle is created via new Bottles(). Line 8 also sends verse(99), which is the action, and then verifies the result with assertEquals.

Running that test produces this error:

TDD tells you to write the simplest code that will pass this test. In this case, your goal is to write only enough code to change the error message. The above error states that the Bottles class does not yet exist, so the first step is to define it, as follows:

If you’re new to TDD, this may seem like a ridiculously small step. Because you wrote the test, you can confidently predict that running it a second time will now produce the following error:

You can change this error message by adding a verse method.

Running the test again produces the following error:

There’s finally sufficient code so that the test fails because the output is not as expected instead of dying because of an exception.

PHPUnit explictly lists what it expected and what it saw. Therefore, you can interpret the above failure as indicating that PHPUnit expected…

but instead got null.

Once you reach this point, it’s easy to make the test pass; just copy the expected output into the verse method:

Although this code passes the test, it clearly doesn’t solve the entire problem. As a matter of fact, writing a second test will break it. While it may seem pointless to write an obviously temporary and transitional bit of code, this is the essence of TDD.

You as the writer of tests know that the verse method must eventually take the value of its argument into account, but you as the writer of code must act in ignorance of this fact. When doing TDD, you toggle between wearing two hats. While in the "writing tests" hat, you keep your eye on the big picture and work your way forward with the overall plan in mind. When in the "writing code" hat, you pretend to know nothing other than the requirements specified by the tests at hand. Thus, although each individual test is correct, until all are written, the code is incomplete.

Now that the first test passes, you must decide what to test next. This next test should do the simplest, most useful thing that proves your existing code to be incorrect. While it may have been difficult to conceive of the first test because the possibilities seem infinite, this next test is often easier because it checks something relative to the first.

Verses 99 through 3 are nearly identical—they differ only in that the number changes within each verse. The test above already checks the high end of this range, and therefore it now makes sense to test the low.

The following test for verse 3 exposes the current verse method to be insufficient:

TDD requires that you pass tests by writing simple code. However, most programming problems have many solutions, and it’s not always clear which one is simplest. For example, the following code passes the current tests by naming the incoming argument and adding a conditional that checks the value of $number and returns the correct string:

At first glance, this code appears to have achieved the ultimate in simplicity. It can produce only the lyrics for verses 99 and 3, and so does the absolute minimum needed to pass the tests.

But consider that it now contains a conditional where none existed before. This may cause you to recall the discussion on Cyclomatic Complexity in Chapter 1. This conditional adds a new execution path through the code, and additional execution paths increase complexity. This code is simple in the sense that it can’t do much, but it does that one small thing in an overly complex way.

Part of the problem is that although the if statement switches on $number, the true and false branches contain many things that don’t vary based on $number. The branches do differ in that one says 99/98, and the other 3/2, but they are the same for all of the other lyrics. This code conflates things that change with things that remain the same, and so forces you to parse strings with your eyes to figure out how $number matters.

If you were to alter the if statement to return only the things that change, the code would look like this:

This code is still very specific to the two existing tests—it can produce the lyrics for verses 99 and 3, and no other. Notice, however, that it now has two parts. The first part (lines 2-6) contains the conditional, and the second (lines 8-12) contains a template that could correctly generate many verses. Lines 2-6 are still specific to the existing tests, but now that you’ve separated the things that change from the things that remain the same, lines 8-12 are generalizable to every verse between 99 and 3.

If you were to continue down the "specific" path, you would progressively add tests for the verses between 97 and 4, each time altering the if statement to add a condition to check for that number. Following this strategy would ultimately result in 97 nearly identical tests and 97 nearly identical verses; each would differ only in the values of the numbers.

The obvious alternative is to instead make the code more general. Because the existing template already works for every verse between 99 and 3, you could change this code to produce those verses by deleting the if statement and altering the template to refer to number, as shown here:

Left to your own devices, your instinct would likely have been to write the code above without bothering with the intermediate steps shown in Listing 2.6: Conditional and Listing 2.7: Sparse Conditional. However, even if you would naturally have started with this more general version, it’s important to understand and be able to articulate the implications of the other implementation.

The difference between the solution that adds a conditional and the solution that adds a variable into a string is that in the first, as the tests get more specific, the code stays equally specific. Every verse has its own personal test and its own individual code; there will never be a time when the code can do anything which is not explicitly tested.

However, in Listing 2.8: Generalized Verse, as the tests get more specific, the code gets more generic. Once the test of verse 3 is written, the code is then generalized to produce lyrics for all verses within the 3-99 range.

Remember that the purpose of this chapter is to quickly get to Shameless Green. With that goal in mind, consider the above solutions and answer this question: Which is simplest?

As previously noted, metrics aren’t everything, but they can certainly be a useful something. In hopes that data will help answer this question, the following chart shows Source Lines Of Code, Cyclomatic Complexity and ABC metrics for the variants, calculated roughly with pencil and paper.

As you can see, as the examples progress, they get shorter, and the Cyclomatic Complexity score improves. The ABC score, on the other hand, gets worse before it gets better. You must, of course, take metrics with a grain of salt, but here they cast a clear vote for replacing duplication with an abstraction as done in Listing 2.8: Generalized Verse.

The next section examines a nearly identical situation where the choice of what to do about duplication is not nearly so clear-cut.

This chapter pulled the lyrics of the "99 Bottles" song out of Bottles and put them into a new BottleVerse class. It then injected an instance of BottleVerse back into Bottles. Extracting BottleVerse reduced the Bottles's responsibilities, making it easier to understand and maintain. Injecting BottleVerse into Bottles loosened the coupling between Bottles and the outside world. Bottles now thinks of itself as being injected with players of the verse template role and will happily collaborate with any newly arriving object as long as that object responds to lyrics(number).

The impetus behind extracting BottleVerse was a new requirement to produce songs with different lyrics, that is, to vary the verse. The refactorings in this chapter satisfied that requirement by following a fundamental strategy of object-oriented design: extracting the BottleVerse class isolated the behavior that the new requirement needed to vary.

While continuing to lean on code smells and refactoring recipes, this chapter introduced the idea of a programming aesthetic. A programming aesthetic is the set of internal heuristics that guide your behavior in times of uncertainty. Vague feelings about the rightness of code become part of your aesthetic once you can eloquently and convincingly use actual words to explain your concerns and proposed improvements. A good programming aesthetic focuses attention on improvements that are likely to prove worthwhile.

This chapter suggested five precepts that belong in everyone’s object-oriented programming aesthetic:

The practical effect of following these precepts is to loosen the coupling between objects. The code to which you would apply them generally already works so adherence might seem optional, and is certainly not free. Complying with these precepts will frequently increase the amount of code and add levels of indirection, at least in the short term. However, these added costs are overwhelmingly offset by the eventual savings accrued as a result of decoupling.

Any application that survives will change. The only thing of which you can be more confident is that you cannot predict where this change will occur. The certainty of change coupled with the uncertainty of that change’s location means that your best programming strategy is to strive to loosen the coupling of all code everywhere from the moment of initial creation.

Therefore, these precepts don’t attempt to guess the future; rather, they leverage against it. Instead of writing code that speculatively imagines a later need for one specific feature, they tell you to loosen the coupling of all code so that you can easily adapt to whatever future arrives.

Uncertainty about the future is not a license to guess; it’s a directive to decouple. Your future will be brighter if you develop a programming aesthetic that drives you to do so.

Tests help prevent errors in code, but to characterize them so simply is a disservice; they offer far more. Good OO is built upon small, interchangeable objects that interact via abstractions. The behavior of each individual object is often quite obvious, but the same cannot be said for the operation of the whole. Tests fill this breach.

Object-oriented applications rely on message sending. The key virtue of messages is that they add indirection. Messages allow the sender to ask for an abstraction and be confident that the receiver will use the appropriate concrete implementation to fulfill the request. Senders are responsible for knowing what they want, receivers, for knowing how to do it. Separating intention from implementation in this way allows you to introduce new variations without altering existing code; simply create a new object that responds to the original message with a different implementation.

When designed with the following features, object-oriented code can interact with new and unanticipated variants without having to change:

Initially, this reliance on abstractions and indirection increases the complexity of code. What OO promises in return is a reduction in the future cost of change. Highly concrete, tightly coupled code will resist tomorrow’s change. Code that depends on loosely coupled abstractions will encourage it.

Because tests need to execute code, they supply early and direct information about inadequate design, and they provide impetus and inspiration for refinements. When tests are difficult to write, require lots of setup, or can’t tell a satisfying story, something is wrong. Listen. Fixing problems now is not only cheaper than fixing them later, but will improve your code, clarify your tests, and make glad your work.