PEARL XVIII : Elucidation on ATDD – Acceptance Test Driven Development
TDD helps software developers produce working, high-quality code that’s maintainable and, most of all, reliable. Customers are rarely, however, interested in buying code. Customers want software that helps them to be more productive, make more money, maintain or improve operational capability, take over a market, and so forth. This is what we need to deliver with our software—functionality to support business function or market needs. Acceptance test-driven development (acceptance TDD) is what helps developers build high-quality software that fulfills the business’s needs as reliably as TDD helps ensure the software’s technical quality.
Acceptance Test Driven Development (ATDD) is a practice in which the whole team collaboratively discusses acceptance criteria, with examples, and then distills them into a set of concrete acceptance tests before development begins. It’s the best way to ensure that we all have the same shared understanding of what it is we’re actually building. It’s also the best way to ensure we have a shared definition of Done. .
Acceptance TDD helps coordinate software projects in a way that helps us deliver exactly what the customer wants when they want it, and that doesn’t let us implement the required functionality only half way.
An essential property of acceptance TDD is that it’s a team activity and a team process.
Acceptance tests are specifications for the desired behavior and functionality of a system. They tell us, for a given user story, how the system handles certain conditions and inputs and with what kinds of outcomes. There are a number of properties that an acceptance test should exhibit;
An important property of acceptance tests is that they use the language of the domain and the customer instead of geek-speak only the programmer understands. This is the fundamental requirement for having the customer involved in the creation of acceptance tests and helps enormously with the job of validating that the tests are correct and sufficient. Scattering too much technical lingo into our tests makes us more vulnerable to having a requirement bug sneak into a production release—because the customer’s eyes glaze over when reading geek-speak and the developers are drawn to the technology rather than the real issue of specifying the right thing.
By using a domain language in specifying our tests, we are also not unnecessarily tied to the implementation, which is useful since we need to be able to refactor our system effectively. By using domain language, the changes we need to make to our existing tests when refactoring are typically non-existent or at most trivial.
Concise, precise, and unambiguous
Largely for the same reasons we write our acceptance tests using the domain’s own language, we want to keep our tests simple and concise. We write each of our acceptance tests to verify a single aspect or scenario relevant to the user story at hand. We keep our tests uncluttered, easy to understand, and easy to translate to executable tests. The less ambiguity involved, the better we are at avoiding mistakes and the working with our tests.
We might write our stories as simple reminders in the form of a bulleted list, or we might opt to spell them out as complete sentences describing the expected behavior. In either case, the goal is to provide just enough information for us to remember the important things we need to discuss and test for, rather than documenting those details beforehand. Card, conversation, confirmation—these are the three Cs that make up a user story. Those same three Cs could be applied to acceptance tests as well.
Yet another common property of acceptance tests is that they might not be implemented (translation: automated) using the same programming language as the system they are testing. Whether this is the case depends on the technologies involved and on the overall architecture of the system under test. For example, some programming languages are easier to inter-operate with than others. Similarly, it is easy to write acceptance tests for a web application through the HTTP protocol with practically any language we want, but it’s often impossible to run acceptance tests for embedded software written in any language other than that of the system itself.
The main reason for choosing a different programming language for implementing acceptance tests than the one we’re using for our production code (and, often, unit tests) is that the needs of acceptance tests are often radically different from the properties of the programming language we use for implementing our system. To give you an example, a particular real-time system might be feasible to implement only with native C code, whereas it would be rather verbose, slow, and error-prone to express tests for the same real-time system in C compared to, for example, a scripting language.
The ideal syntax for expressing our acceptance tests could be a declarative, tabular structure such as a spreadsheet, or it could be something closer to a sequence of higher-level actions written in plain English. If we want to have our customer collaborate with developers on our acceptance tests, a full-blown programming language such as Java, C/C++, or C# is likely not an option. “Best tool for the job” means more than technically best, because the programmer’s job description also includes collaborating with the customer.
The acceptance TDD cycle
In its simplest form, the process of acceptance test-driven development can be expressed as the simple cycle illustrated by figure 1
Figure 1. The acceptance TDD cycle
This cycle continues throughout the iteration as long as we have more stories to implement, starting over again from picking a user story; then writing tests for the chosen story, then turning those tests into automated, executable tests; and finally implementing the functionality to make our acceptance tests pass.
In practice, of course, things aren’t always that simple. We might not yet have user stories, the stories might be ambiguous or even contradictory, the stories might not have been prioritized, the stories might have dependencies that affect their scheduling, and so on.
Step 1: Pick a user story
The first step is to decide which story to work on next. Not always an easy job; but, fortunately, most of the time we’ll already have some relative priorities in place for all the stories in our iteration’s work backlog. Assuming that we have such priorities, the simplest way to go is to always pick the story that’s on top of the stack—that is, the story that’s considered the most important of those remaining. Again, sometimes, it’s not that simple.
Generally speaking, the stories are coming from the various planning meetings held throughout the project where the customer informally describes new features, providing examples to illustrate how the system should work in each situation. In those meetings, the developers and testers typically ask questions about the features, making them a medium for intense learning and discussion. Some of that information gets documented on a story card (whether virtual or physical), and some of it remains as tacit knowledge. In those same planning meetings, the customer prioritizes the stack of user stories by their business value (including business risk) and technical risk (as estimated by the team).
There are times when the highest-priority story requires skills that we don’t possess, or we consider not having enough of. In those situations, we might want to skip to the next task to see whether it makes more sense for us to work on it. Teams that have adopted pair programming don’t suffer from this problem as often. When working in pairs, even the most cross-functional team can usually accommodate by adjusting their current pairs in a way that frees the necessary skills for picking the highest priority task from the pile.
The least qualified person
The traditional way of dividing work on a team is for everyone to do what they do best. It’s intuitive. It’s safe. But it might not be the best way of completing the task. Arlo Belshee presented an experience report at the Agile 2005 conference, where he described how his company had started consciously tweaking the way they work and measuring what works and what doesn’t. Among their findings about stuff that worked was a practice of giving tasks to the least qualified person.
There can be more issues to deal with regarding picking user stories, but most of the time the solution comes easily through judicious application of common sense. For now, let’s move on to the second step in our process: writing tests for the story we’ve just picked.
Step 2: Write tests for a story
With a story card in hand (or onscreen if you’ve opted for managing your stories online), our next step is to write tests for the story.
The first thing to do is, of course, get together with the customer. In practice, this means having a team member sit down with the customer (they’re the one who should own the tests, remember?) and start sketching out a list of tests for the story in question.
As usual, there are personal preferences for how to go about doing this, but current preference is to quickly scram out a list of rough scenarios or aspects of the story we want to test in order to say that the feature has been implemented correctly. There’s time to elaborate on those rough scenarios later on when we’re implementing the story or implementing the acceptance tests. At this time, however, we’re only talking about coming up with a bulleted list of things we need to test—things that have to work in order for us to claim the story is done.
Especially in projects that have been going on for a while already, the customer and the development team probably have some kind of an idea of what’s going to get scheduled into the next iteration in the upcoming planning meeting. In such projects, the customer and the team have probably spent some time during the previous iteration sketching acceptance tests for the features most likely to get picked in the next iteration’s planning session. This means that we might be writing acceptance tests for stories that we’re not going to implement until maybe a couple of weeks from now. We also might think of missing tests during implementation, for example, so this test-writing might happen pretty much at any point in time between writing the user story and the moment when the customer accepts the story as completed.
Once we have such a rough list, we start elaborating the tests, adding more detail and discussing about how this and that should work, whether there are any specifics about the user interface the customer would like to dictate, and so forth. Depending on the type of feature, the tests might be a set of interaction sequences or flows, or they might be a set of inputs and expected outputs. Often, especially with flow-style tests, the tests specify some kind of a starting state, a context the test assumes is part of the system.
Other than the level of detail and the sequence in which we work to add that detail, there’s a question of when—or whether—to start writing the tests into an executable format. Witness step 3 in our process: automating the tests.
Step 3: Automate the tests
The next step once we’ve got acceptance tests written down on the back of a story card, on a whiteboard, in some electronic format, or on pink napkins, is to turn those tests into something we can execute automatically and get back a simple pass-or-fail result. Whereas we’ve called the previous step writing tests, we might call this step implementing or automating those tests.
In an attempt to avoid potential confusion about how the executable acceptance tests differ from the acceptance tests
We might turn Acceptance tests into an executable format by using a variety of approaches and tools. The most popular category of tools these days seems to be what we calltable-based tools. Their premise is that the tabular format of tables, rows, and columns makes it easy for us to specify our tests in a way that’s both human and machine readable. Figure 2 presents an example of how we might draft an executable test for the first test “Valid account number”.
Figure 2. Example of an executable test, sketched on a piece of paper
In figure 2, we’ve outlined the steps we’re going to execute as part of our executable test in order to verify that the case of an incoming support call with a valid account number is handled as expected, displaying the customer’s information onscreen. Our test is already expressed in a format that’s easy to turn into a tabular table format using our tool of choice—for example, something that eats HTML tables and translates their content into a sequence of method invocations to Java code according to some documented rules.
The inevitable fact is that most of the time, there is not such a tool available that would understand our domain language tests in our table format and be able to wire those tests into calls to the system under test. In practice, we’ll have to do that wiring ourselves anyway—most likely the developers or testers will do so using a programming language.
To summarize this duality of turning acceptance tests into executable tests, we’re dealing with expressing the tests in a format that’s both human and machine readable and with writing the plumbing code to connect those tests to the system under test.
The example in figure 2 is a flow-style test, based on a sequence of actions and parameters for those actions. This is not the only style at our disposal, however. A declarative approach to expressing the desired functionality or business rule can often yield more compact and more expressive tests than what’s possible with flow-style tests.
Yet our goal should—once again—be to keep our tests as simple and to the point as possible, ideally speaking in terms of what we’re doing instead of how we’re doing it.
With regard to writing things down , there are variations on how different teams do this. Some start writing the tests right away into electronic format using a word processor; some even go so far as to write them directly in an executable syntax. Some teams run their tests as early as during the initial authoring session. Some people, prefer to work on the tests alongside the customer using a physical medium, leaving the running of the executable tests for a later time. For example, agilists like to sketch the executable tests on a whiteboard or a piece of paper first, and pick up the computerized tools only when they got something relatively sure won’t need to be changed right away.
The benefit is that we’re less likely to fall prey to the technology—Agilsts noticed that tools often steal too much focus from the topic, which we don’t want. Using software also has this strange effect of the artifacts being worked on somehow seeming more formal, more final, and thus needing more polishing up. All that costs time and money, keeping us from the important work.
In projects where the customer’s availability is the bottleneck, especially in the beginning of an iteration (and this is the case more often than not), it makes a lot of sense to have a team member do the possibly laborious or uninteresting translation step on their own rather than keep the customer from working on elaborating tests for other stories. The downside to having the team member formulate the executable syntax alone is that the customer might feel less ownership in the acceptance tests in general—after all, it’s not the exact same piece they were working on. Furthermore, depending on the chosen test-automation tool and its syntax, the customer might even have difficulty reading the acceptance tests once they’ve been shoved into the executable format dictated by the tool.
let’s consider a case where our test-automation tool is a framework for which we express our tests in a simple but powerful scripting language such as Ruby. Figure 3 highlights the issue with the customer likely not being as capable of feeling ownership of the implemented acceptance test compared to the sketch, which they have participated in writing. Although the executable snippet of Ruby code certainly reads nicely to a programmer, it’s not so trivial for a non-technical person to relate to.
Figure 3. Contrast between a sketch an actual, implemented executable acceptance test
Another aspect to take into consideration is whether we should make all tests executable to start with or whether we should automate one test at a time as we progress with the implementation. Some teams—and this is largely dependent on the level of certainty regarding the requirements—do fine by automating all known tests for a given story up front before moving on to implementing the story.
Some teams prefer moving in baby steps like they do in regular test-driven development, implementing one test, implementing the respective slice of the story, implementing another test, and so forth. The downside to automating all tests up front is, of course, that we’re risking more unfinished work—inventory, if you will—than we would be if we’d implemented one slice at a time. Agilists preference is strongly on the side of implementing acceptance tests one at a time rather than try getting them all done in one big burst. It should be mentioned, though, that elaborating acceptance tests toward their executable form during planning sessions could help a team understand the complexity of the story better and, thus, aid in making better estimates.
Many of the decisions regarding physical versus electronic medium, translating to executable syntax together or not, and so forth also depend to a large degree on the people. Some customers have no trouble working on the tests directly in the executable format (especially if the tool supports developing a domain-specific language). Some customers don’t have trouble identifying with tests that have been translated from their writing. As in so many aspects of software development, it depends.
Regardless of our choice of how many tests to automate at a time, after finishing this step of the cycle we have at least one acceptance test turned into an executable format; and before we proceed to implementing the functionality in question, we will have also written the necessary plumbing code for letting the test-automation tool know what those funny words mean in terms of technology. That is, we will have identified what the system should do when we say “select a transaction” or “place a call”—in terms of the programming API or other interface exposed by the system under test.
To put it another way, once we’ve gotten this far, we have an acceptance test that we can execute and that tells us that the specified functionality is missing. The next step is naturally to make that test pass—that is, implement the functionality to satisfy the failing test.
Step 4: Implement the functionality
Next on our plate is to come up with the functionality that makes our newly minted acceptance test(s) pass. Acceptance test-driven development doesn’t say how we should implement the functionality; but, needless to say, it is generally considered best practice among practitioners of acceptance TDD to do the implementation using test-driven development.
In general, a given story represents a piece of customer-valued functionality that is split—by the developers—into a set of tasks required for creating that functionality. It is these tasks that the developer then proceeds to tackle using whatever tools necessary, including TDD. When a given task is completed, the developer moves on to the next task, and so forth, until the story is completed—which is indicated by the acceptance tests executing successfully.
In practice, this process means plenty of small iterations within iterations. Figure 4 visualizes this transition to and from test-driven development inside the acceptance TDD process.
Figure 4. The relationship between test-driven development and acceptance test-driven development
As we can see, the fourth step of the acceptance test-driven development cycle, implementing the necessary functionality to fix a failing acceptance test, can be expanded into a sequence of smaller TDD cycles of test-code-refactor, building up the missing functionality in a piecemeal fashion until the acceptance test passes.
While the developer is working on a story, frequently consulting with the customer on how this and that ought to work, there will undoubtedly be occasions when the developer comes up with a scenario—a test—that the system should probably handle in addition to the customer/developer writing those things down. Being rational creatures, we add those acceptance tests to our list, perhaps after asking the customer what they think of the test. After all, they might not assign as much value to the given aspect or functionality of the story as we the developers might.
At some point, we’ve iterated through all the tasks and all the tests we’ve identified for the story, and the acceptance tests are happily passing. At this point, depending on whether we opted for automating all tests up front or automating them just in time, we either go back to Step 3 to automate another test or to Step 1 to pick a brand-new story to work on.
. Getting acceptance tests passing is intensive work.
Acceptance TDD inside an iteration
A healthy iteration consists mostly of hard work. Spend too much time in meetings or planning ahead, and you’re soon behind the iteration schedule and need to de-scope . Given a clear goal for the iteration, good user stories, and access to someone to answer our questions, most of the iteration should be spent in small cycles of a few hours to a couple of days writing acceptance tests, collaborating with the customer where necessary, making the tests executable, and implementing the missing functionality with test-driven development.
As such, the four-step acceptance test-driven development cycle of picking a story, writing tests for the story, implementing the tests, and implementing the story is only a fraction of the larger continuum of a whole iteration made of multiple—even up to dozens—of user stories, depending on the size of your team and the size of your stories. In order to gain understanding of how the small four-step cycle for a single user story fits into the iteration, we’re going to touch the zoom dial and see what an iteration might look like on a time line with the acceptance TDD–related activities scattered over the duration of a single iteration.
Figure 5 is an attempt to describe what such a time line might look like for a single iteration with nine user stories to implement. Each of the bars represents a single user story moving through the steps of writing acceptance tests, implementing acceptance tests, and implementing the story itself. In practice, there could (and probably would) be more iterations within each story, because we generally don’t write and implement all acceptance tests in one go but rather proceed through tests one by one.
Figure 5. Putting acceptance test-driven development on time line
Notice how the stories get completed almost from the beginning of the iteration? That’s the secret ingredient that acceptance TDD packs to provide indication of real progress. Our two imaginary developers (or pairs of developers and/or testers, if we’re pair programming) start working on the next-highest priority story as soon as they’re done with their current story. The developers don’t begin working on a new story before the current story is done. Thus, there are always two user stories getting worked on, and functionality gets completed throughout the iteration.
So, if the iteration doesn’t include writing the user stories, where are they coming from? As you may know if you’re familiar with agile methods, there is usually some kind of a planning meeting in the beginning of the iteration where the customer decides which stories get implemented in that iteration and which stories are left in the stack for the future. Because we’re scheduling the stories in that meeting, clearly we’ll have to have those stories written before the meeting, no?
That’s where continuous planning comes into the picture.
Although an iteration should ideally be an autonomous, closed system that includes everything necessary to meet the iteration’s goal, it is often necessary—and useful—to prepare for the next iteration during the previous one by allocating some amount of time for pre-iteration planning activities. Suggestions regarding the time we should allocate for this continuous planning range from 10–15% of the team’s total time available during the iteration. As usual, it’s good to start with something that has worked for others and, once we’ve got some experience doing things that way, begin zeroing in on a number that seems to work best in our particular context.
In practice, these pre-iteration planning activities might involve going through the backlog of user stories, identifying stories that are most likely to get scheduled for the next iteration, identifying stories that have been rendered obsolete, and so forth. This ongoing pre-iteration planning is also the context in which we carry out the writing of user stories and, to some extent, the writing of the first acceptance tests. The rationale here is to be prepared for the next iteration’s beginning when the backlog of stories is put on the table. At that point, the better we know our backlog, the more smoothly the planning session goes, and the faster we get back to work, crunching out valuable functionality for our customer.
By writing, estimating, splitting if necessary, and prioritizing user stories before the planning meeting, we ensure quick and productive planning meetings and are able to get back to delivering valuable features sooner.
It would be nice if we had all acceptance tests implemented (and failing) before we start implementing the production code. That is often not a realistic scenario, however, because tests require effort as well—they don’t just appear from thin air—and investing our time in implementing the complete set of acceptance tests up front doesn’t make any more sense than big up-front design does in the larger scale. It is much more efficient to implement acceptance tests as we go, user story by user story.
Teams that have dedicated testing personnel can have the testing engineers work together with the customer to make acceptance tests executable while developers start implementing the functionality for the stories. A hazard is that most teams, however, are much more homogeneous in this regard and participate in writing and implementing acceptance tests together, with nobody designated as “the acceptance test guy.”
The process is largely dependent on the availability of the customer and the test and software engineers. If your customer is only onsite for a few days in the beginning of each iteration, you probably need to do some trade-offs in order to make the most out of those few days and defer work that can be deferred until after the customer is no longer available. Similarly, somebody has to write code, and it’s likely not the customer who’ll do that; software and test engineers need to be involved at some point.
We start from those stories we’ll be working on first, of course, and implement the user story in parallel with automating the acceptance tests that we’ll use to verify our work. And, if at all possible, we avoid having the same person implement the tests and the production code in order to minimize our risk of human nature playing its tricks on us.
Again, we want to keep an eye on putting too much up-front effort in automating our acceptance tests—we might end up with a huge bunch of tests but no working software. It’s much better to proceed in small steps, delivering one story at a time. No matter how valuable our acceptance tests are to us, their value to the customer is negligible without the associated functionality.
The mid-iteration sanity check
Agilists like to have an informal sanity check in the middle of an iteration. At that point, we should have approximately half of the stories scheduled for the iteration running and passing. This might not be the case for the first iteration, due to having to build up more infrastructure than in later iterations; but, especially as we get better at estimating our stories, it should always be in the remote vicinity of having 50% of the stories passing their tests.
Of course, we’ll be tracking story completion throughout the iteration. Sometimes we realize early on that our estimated burn rate was clearly off, and we must adjust the backlog immediately and accordingly. By the middle of an iteration, however, we should generally be pretty close to having half the stories for the iteration completed. If not, the chances are that there’s more work to do than the team’s capacity can sustain, or the stories are too big compared to the iteration length.
A story’s burn-down rate is constantly more accurate a source of prediction than an inherently optimistic software developer. If it looks like we’re not going to live up to our planned iteration content, we decrease our load.
Decreasing the load
When it looks like we’re running out of time, we decrease the load. We don’t work harder (or smarter). We’re way past that illusion. We don’t want to sacrifice quality, because producing good quality guarantees the sustainability of our productivity, whereas bad quality only creates more rework and grinds our progress to a halt. We also don’t want to have our developers burn out from working overtime, especially when we know that working overtime doesn’t make any difference in the long run. Instead, we adjust the one thing we can: the iteration’s scope—to reality. In general, there are three ways to do that: swap, drop, and split. Tom DeMarco and Timothy Lister have done a great favor to our industry with their best-selling books Slack (DeMarco; Broadway, 2001) and Peopleware (DeMarco, Lister; Dorset House, 1999), which explain how overtime reduces productivity.
Swapping stories is simple. We trade one story for another, smaller one, thereby decreasing our workload. Again, we must consult the customer in order to assure that we still have the best possible content for the current iteration, given our best knowledge of how much work we can complete.
Dropping user stories is almost as straightforward as swapping them. “This low-priority story right here, we won’t do in this iteration. We’ll put it back into the product backlog.” But dropping the lowest-priority story might not always be the best option, considering the overall value delivered by the iteration—that particular story might be of low priority in itself, but it might also be part of a bigger whole that our customer cares about. We don’t want to optimize locally. Instead, we want to make sure that what we deliver in the end of the iteration is a cohesive whole that makes sense and can stand on its own.
The third way to decrease our load, splitting, is a bit trickier compared to dropping and swapping
How do we split a story we already tried hard to keep as small as possible during the initial planning game? In general, we can split stories by function or by detail (or both). Consider a story such as “As a regular user of the online banking application, I want to optionally select the recipient information for a bank transfer from a list of most frequently and recently used accounts based on my history so that I don’t have to type in the details for the recipients every time.”
Splitting this story by function could mean dividing the story into “…from a list of recently used accounts” and “…from a list of most frequently used accounts.” Plus, depending on what the customer means by “most frequently and recently used,” we might end up adding another story along the lines of “…from a weighted list of most frequently and recently used accounts” where the weighted list uses an algorithm specified by the customer. Having these multiple smaller stories, we could then start by implementing a subset of the original, large story’s functionality and then add to it by implementing the other slices, building on what we have implemented for the earlier stories.
Splitting it by detail could result in separate stories for remembering only the account numbers, then also the recipient names, then the VAT numbers, and so forth. The usefulness of this approach is greatly dependent on the distribution of the overall effort between the details—if most of the work is in building the common infrastructure rather than in adding support for one more detail, then splitting by function might be a better option. On the other hand, if a significant part of the effort is in, for example, manually adding stuff to various places in the code base to support one new persistent field, splitting by detail might make sense.
Regardless of the chosen strategy, the most important thing to keep in mind is that, after the splitting, the resulting user stories should still represent something that makes sense—something valuable—to the customer.