This challenge is the tricky problem of explaining a missing value.
What’s the best way to describe a Given for a value that’s not supposed to be there? For example, starting the specification from a point of a user who does not yet exist in a database, or a product that’s not been set up. “Given a user that doesn’t exist” sounds vague, but “Given user John that does not exist” sounds silly – how can we talk about John if he doesn’t exist yet?
Key examples should ideally be specific enough to avoid ambiguity, but it’s difficult to be specific about something that does not exist. Such problematic Given-When-Then scenarios are usually overly generic, and do not really help with shared understanding or testing. Here’s a typical trivial example that should ensure unique accounts:
Given user John does not exist
When John tries to register
Then the registration is successful
Given user John exists
When John tries to register
Then the registration fails
This set of scenarios seems simple but it can hide many important constraints, and provide a false sense of shared understanding. A team might think they nailed down a feature, but develop something full of bugs. Scenarios such as these leave too much for later exploratory testing, overloading the team to discover seemingly unexpected problems that could have easily been predicted and documented with better examples.
As a trivial counter-example, consider what should happen if John tries to register with an email already assigned to a different user. For a more complex edge case, consider what should happen if two Johns tried to register at the same time. And what is “John” anyway in this case? Is it a personal name or a username? If it’s a personal name, should we really prevent two people with the same name from opening two different accounts?
Oversimplified examples often lead to overcomplicated test automation. Proving that a user successfully registered usually requires accessing a database, which means that the related tests will be slow and brittle. Accessing external systems is a huge performance penalty compared to in-process tests. Asynchronous networking introduces a whole set of technical edge cases that have nothing to do with user registration business rules.
Data persistence leads to issues around reproducibility. John might not exist in the database the first time you run a test, but it surely will be there after that test completes. To make such a test repeatable, you’ll either need to introduce complex test set-up or clean-up procedures or make the examples even more generic by introducing placeholders for random or unique values.
Database resources are usually difficult to set up from scratch, so team members and test automation systems sometimes share testing databases, which can lead to issues around reproducibility and isolation. If two people execute tests against the same database at the same time, one might add data that the other does not expect, causing tests to fail intermittently without a good explanation.
Last time, we started a community challenge to suggest improvements to a badly worded Given-When-Then feature file, specifically addressing the problem of setting up something that’s not supposed to exist. This article contains an analysis of the responses, some tips on a more general approach to solving similar problems, and a new challenge for this week.
I’ve outlined some of the most common problems with describing something that doesn’t exist in the original challenge, so I will not repeat them here – check the linked post for a refresher.
We received lots of excellent responses to the first challenge, from people in sixteen different countries. Alister Scott even wrote a wonderful blog post about his proposal.
Although the problems with describing something that doesn’t exist sound like technical challenges that are purely related to testing, they are often just symptoms of an unexplored model. It’s possible to solve or avoid most of them by exploring the model differently, and ensuring shared understanding between team members and stakeholders. Concrete examples are a great way to do that, and many solutions tried to make things more concrete by introducing different properties of users, or talking about users before they register in a different way.
For example, several suggestions involved introducing a specific qualifier, such as an “unregistered user”. This helps to differentiate between two different types of entities: one that exists before registration, and another that exists after. Similarly, some solutions tried to move away from an ambiguous user identifiers, such as “john”, to more concrete data available even users aren’t registered. For example, it’s possible to consider user emails even before they sign up to our system. The following suggestion, which also came in anonymously, illustrates this nicely:
Given user firstname.lastname@example.org tries to create an account When there is already an account for email@example.com Then registration is not possible
All these attempts are going in the right direction, but they are still just workarounds for a problem and they aren’t solving it fully. On a conceptual level, a great way to approach similar problems is another anonymous suggestion, “Describe the state of the system without the value”. Instead of trying to describe something that is not supposed to exist, we can describe what else exists instead.
There are two ways of approaching this suggestion practically. The first, suggested by Mathieu Roseboom, is to look outside the system. Mathieu wrote “I’d emphasize that John is a person, and not a user.” Rather than talking about unregistered users that do not exist, let’s talk about people that do exist. The suggestions that tried to use emails to describe unregistered users lean in that direction, without fully benefiting from it. Thinking about person that exists outside of our system might leads us to discover about some other attributes, which could be important for the current feature.
The second way to specify something that exists instead of talking about vague non-existing entities is to describe all the other registered users. For example, one response suggested splitting the problem into two cases:
Given list of users is empty ... Given a "normal" database
Another reader suggested setting up a separate feature file, that just creates the relevant users, and then technically ensuring that all other feature file tests run after it.
Feature: Setup users Scenario: Register users The following users are registered: | name | email | | Rick | firstname.lastname@example.org |
David Wardlaw suggested starting the scenario with:
Given the following user is already registered | First Name | Last Name | Email Address | | John | Smith | email@example.com | And a new user wants to register with a first name of <First Name>
These three ideas illustrate different levels of visibility that feature files could offer. In the first option, readers will need to know what the “normal” database contains, but the solution is quite easy to reuse across different scenarios. On the other hand, if someone modifies the “normal” database for some unrelated reason, tests might start to break unexpectedly.
The second solution optimises performance because it inserts records only once, and provides better visibility to readers about the assumed state. The downside of this idea is that it imposes a very specific order for executing tests. This feature file must run before everything else, otherwise things start weirdly misbehaving. I usually try to avoid imposing a specific order of test executions, since then we can’t just run a single test when needed. This approach also suffers from implied shared state. If someone updates that central setup file for some unrelated test case, our tests may magically start to fail.
The third suggestion provides full visibility to a reader, and it does not impose any shared state that could cause problems later when things change. This approach might get a bit wordy if we need to set up a lot of users, but there are ways around that as well.
The key problem the first and second solution are trying to solve is the complexity and performance of working with a real database. Databases are slow and difficult to control compared to simple in-memory objects, so tests involving a real database often have to compensate for those downsides somehow. Between the three ideas, I would go with the third one unless there is some very specific performance constraint we want to solve. I promise to come back to this next week, but since we’re not solving database performance in this challenge, let’s just postpone that discussion.
So which approach should we choose? For that, I’d like to explore this problem on a more conceptual level, and explain several techniques which you can use with other similar situations.
In order to reason about the existence of an item (such as a user or a product), we first need to answer a key probing question: where? Where should that item exist or not exist? Defining existence is, at core, reasoning about the state of system data. Figuring where the data resides is critical for three reasons:
Team members closer to implementation work, such as developers or testers, often think about entities as system data. They will intuitively understand that an entity always exists in a specific location, but they might not be able to share their assumptions about this easily with other team members. People not used to working with system state, such as business representatives, might struggle with that idea. They tend to think about existence in a more absolute way. A helpful probing technique to start the conversation about this concept is to ask an extreme question.
These questions might sounds silly, but try them out and you’ll quickly see their value as conversation starters. Extreme questions often result with a hard “No”, and get people to start thinking about multiple contexts, a timeline or a scale. You will start identifying where an entity exists directly before it appears in your context, and what related information you can actually reason about at that point. Mathieu Roseboom’s suggestion to talk about a “person” is one potential outcome of such discussions.
The first set of answers to an extreme question is usually an overly generic statement. For example, the user might exist as a living person, but they don’t yet exist in “the system” or in “the database”. That’s a good starting point for the discussion, since we can now consider what’s known about an entity in different contexts. We can explore the domain much more easily.
A non-registered user likely exists as a living person, so we can talk about their personal name. They do not yet have a username in our system, so we can’t talk about it yet. They likely have an email, which is why so many responses to the challenge focused on that attribute. They might also have a preferred username in mind, which has nothing to do with them existing or not existing in our system. The When part can then become clearer:
When John Michaels attempts to register with the preferred username "john"
A good technique to continue the discussion about something generic such as “system” or “database” is to provide a meaningful domain name. In the Domain-Driven-Design terminology, domain representations of entity stores are called Repositories. When developers hear the word “repository”, they often think about a common technical pattern that involves specific technical operations. Ignore that for the moment – that’s implementation detail. A repository is a meaningful first-order domain concept that encapsulates storing and retrieving data. Tell the rest of the team to imagine it as a box where you keep the users. If they can’t get over that, then think of a different name. I’ll keep using “repository” in this post.
Instead of a user existing or not existing in an absolute sense, we can potentially talk about the repository:
Given John does not exist in the user repository
Even better, let’s make the user repository the subject of the sentence. The repository exists regardless of the users:
Given a user repository that does not include a user called John
The real precondition here is that a user repository exists, with some specific constraints about its contents. There’s no more chicken-and-egg issue. We identified an important first-order domain concept that we can reason about, regardless of a specific user.
Instead of just saying “does not include a user”, which is still a bit vague, we can now start capturing the constraints of the user repository in a more specific way, using an approach very similar to what David Wardlaw suggested. Here’s how I’d start writing it:
Given a user repository with the following users | username | personal name | | mike | Mike Smith | | steve123 | Steve James | When John Michaels attempts to register with the username "john" Then the user repository should contain the following users: | username | personal name | | mike | Mike Smith | | steve123 | Steve James | | john | John Michaels |
We now have a simple but concrete example. We’re not done yet, this is still just a good conversation starter. A good technique to continue the discussion is Simple-Counter-Key:
The first step is to add at least one simple example, which could lead to a different outcome:
When Steve James attempts to register with the username "steve123"
The outcome in this case might be obvious to everyone, but vary the data a bit. What happens if another “Steve James” attempts to register with the username “steveo”?
This can lead to an interesting conversation around the meaning of “uniqueness” and “existence”. Is the purpose of this feature to prevent the same person from creating multiple accounts in the system? If so, we should probably stop Steve from registering again, even with a different username. But if two different people called Steve James try to register, we should not prevent them. We need some other way of determining uniqueness, and personal names are obviously not good enough. Still, we might care about personal names in this scenario to ensure that we’re capturing them correctly in the repository. Are there any other attributes of a person that we should care about? This brings us back to the emails, suggested by several readers responding to the challenge. But how do we know that emails are the right thing to capture?
We can now start discussing the meaning of “unique” as relating to the person, and if emails are the right way to approach it. Some systems need to be lax about this, so they might let people register even with the same email. For example, a multi-tenant cloud app might want to allow opening sub-accounts with the same admin email to provide centralised billing. Those users will genuinely be different people, sharing a common email inbox. Many online systems today want to prevent abuse by validating emails, so they insist that every user has a unique email. Some systems need to be more strict, and they might enforce a unique mobile phone number. For government or banking purposes, even that’s not enough, and we might want to enforce unique date of birth, social security number, passport number or some other government-issued identifier. Different contexts will have different rules around this. Exposing those rules often leads to much better shared understanding, and also points to the attributes which should be captured for the key examples.
Discussing simple examples and counterexamples allows us to probe further into the model. What does it actually mean for a user to exist in our context? Does that mean the user repository has a record matching the specified email, or proposed username, or something else? How do we know that two users are not the same person? To keep things simple, let’s limit our case to just checking that the emails are different. We’ll say that a user exists in the system even if the proposed username is not taken, but we have a different username matching the same email. The simple scenario now evolves to capture the email as well:
Scenario: Users with a unique email should register successfully Given a user repository with the following users: | username | personal name | email | | mike | Mike Smith | firstname.lastname@example.org | | steve123 | Steve James | email@example.com | When John Michaels attempts to register with the username "john" and email "firstname.lastname@example.org" Then the user repository should contain the following users: | username | personal name | email | | mike | Mike Smith | email@example.com | | steve123 | Steve James | firstname.lastname@example.org | | john | John Michaels | email@example.com |
We could add three or four more examples checking for duplicated emails and usernames in the same way, but this will very quickly become difficult to read.
If you’ve not yet read Alister Scott’s blog post about this challenge, now would be a good time to do it. He argues against misusing Given-When-Then to capture similar examples with small variances over and over again, suggesting that tables would be a better match for this case.
Tables are indeed a much better way of capturing related examples, but they can sometimes make it difficult to understand the purpose of a specific example from the group. Alister solves this nicely by pointing out differences between examples in an additional field, called Result. This field contains the explanation why a case failed, and what should have happened instead. For example, one of the results in his blog post is “Registration Unsuccessful – Reset password page displayed with email prefilled”. This points to another interesting thing we might want to check. Knowing that registration failed may not be enough – we need to know that it failed for the right reasons.
I agree with Alister that misusing Given-When-Then to copy and paste examples is a waste of time (and screen space). Tables are a good solution, and most Given-When-Then tools support listing tables of examples with scenario outlines. Scenario outlines use placeholders marked with
<> inside a Given-When-Then scenario, followed by a table of examples with placeholder values. Here is how we could show two failure examples in the same structure:
Scenario Outline: Users with existing emails or usernames should be rejected Given a user repository with the following users: | username | personal name | email | | mike | Mike Smith | firstname.lastname@example.org | | steve123 | Steve James | email@example.com | When the Tony James tries to register with <requested username> and <email> Then the registration should fail with <registration message> and user repository should contain only the following users: | username | personal name | email | | mike | Mike Smith | firstname.lastname@example.org | | steve123 | Steve James | email@example.com | Examples: | requested username | email | registration message | | steve123 | firstname.lastname@example.org | Username taken | | steve456 | email@example.com | Email already registered |
Note that the When and Then statements contain placeholders relating to the table of examples, which is shown at the bottom.
Evolving Alister’s “result” idea, the registration message helps to explain what’s going on, but it’s also an important domain concept. Now that we identified it, we can talk about what that should contain. For a closely controlled system, we might want to immediately remind an existing user about his username, so he can sign in easily instead of registering or calling support staff. For an open public system, we might want to go in the other direction and mask the fact that an email is registered, and prevent potential malicious actors from inspecting accounts by trying to sign up.
Showing a comment in the table of examples is quite a common trick to explain the purpose of individual examples. Some people prefer to use the first column for that. I tend to use the first column if it’s a generic comment or name of the test, and does not play any role in the testable outputs. If we actually want to check it against some system output (in this case the registration message), I prefer to list inputs on the left and outputs on the right side of the table.
The technique I mentioned earlier is Simple-Counter-Boundary. We still need to do the third step. So far we identified a few simple examples and some counter-examples. That’s a great start because it gives us the structure for discovering boundaries. Inspect each individual property in the Given section and probe with further examples.
What if someone tries to register with the proposed username “Steve123”? How about the email “Steve@Yahoo.com”? An email written in a different case is still the same, so the registration should probably fail. Some systems (notably Amazon AWS Cognito) by default do not enforce case-sensitive username checks, so two people could register with just a small difference in spelling. This is a support and maintenance nightmare, so let’s make sure we prevent it.
The nice thing about a scenario outline is that it’s very easy to add more examples. We can just append the following two cases to the table:
| Steve123 | firstname.lastname@example.org | Username taken | | steve456 | Steve@Yahoo.com | Email already registered |
If we really want to avoid support problems with accounts that are too similar, perhaps we should also prevent people from registering usernames that only differ from existing accounts in interpunction symbols.
| steve.456 | Steve3@Yahoo.com | Username taken | | steve_456 | Steve3@Yahoo.com | Username taken |
Discussing boundaries often leads to more examples, and identifying hidden domain rules. And again, some systems will have different rules. Expose an example, discuss it, then decide if you want to keep it or not. Perhaps the interpunction checking in usernames is too much or too complicated, so we can skip that for now. We could monitor if it becomes a problem, and add it to a later iteration if necessary. In most teams, it’s up to developers and testers to offer this for discussion, and for business representatives to decide on priority and scoping.
Even if the username interpunction examples for usernames end up out of scope, thinking about this them can lead you to consider similar boundaries for other input fields. Should we care about interpunction symbols in emails?
For example, Gmail allows users to put a dot anywhere in the email, or follow the email with a plus sign and a label, so “email@example.com” and “firstname.lastname@example.org” are actually the same physical account, as well as “email@example.com”. Popular email systems often have aliases, so “firstname.lastname@example.org” is the same as “email@example.com”. How much do we care about preventing duplicated emails? Should we try to fight against commonly known cases such as that, or just ignore them? These are very specific domain questions, and the answers will depend on the risks you are trying to control. For a complex set of rules, there may be many more examples, and this table might become too big to read.
This brings us to the next challenge. Let’s say that we actually do want to be a bit more strict about preventing the same person from registering twice, so all those Gmail hacks need to be handled. Plus the support managers came back and said that they liked the idea of avoiding problems with usernames that are too similar. This will lead to dozens, if not hundreds of examples. Alister also looks at examples around password complexity, which is another common aspect of registration that we might want to add to the registration feature. Putting it all into a single scenario outline is definitely not a good idea. How would you better structure such a big list of examples, so it’s easy to understand and maintain?