Two styles of step writing in Gherkin sparked a discussion about Cucumber practices here in the office. Michael has been doing a lot of research, reflection, and has been taking a lot of good notes from The Cucumber Book. We this in mind, we recently sat down as a group to attempt to define some Gaslight specific Cucumber conventions and best practices.
Here are examples of the two styles that sparked the discussion:
Scenario: User has questions Given I belong to a group "Sales Reps" And the group "Sales Reps" has a question When I login Then I should be prompted with the question for "Sales Reps"
Scenario: User has questions Given I belong to a group with a question When I login Then I should be prompted with the question
The assumption is that each style better serves a certain contingent. The first contingent being developers of the system. The other being the non-technical clients or owners of the system.
The second example is customer facing. It’s in plain english. It’s easy to read and reason about. However it prohibits reuse, and the steps cannot exist in isolation. Implementers will need to introduce instance variables in order to tie together the Given and Then steps. As a developer this is hard, our tendency is eliminate coupling and always promote code reuse.
It follows from software design practice that steps written in this way create more work for the developer. However, I the more I think about this, the more I start to disagree.
Why are we doing this?
The goal of any kind of acceptance testing should be to keep stake-holders engaged with the project. It’s a way for non-technical people to communicate the functionality of the system in a practical way. It creates a document that describes exactly what the system should do, and simultaneously verifies that it does it.
Stake-holder involvement is the most important factor in a software project’s success. Therefore, every effort should be make to write Gherkin in a way that communicates with the stake-holder and keeps them engaged. Don’t bore the stakeholders!
If there were no benefit to be gained from stake-holder involvement, there would be absolutely no need for native Ruby speakers to write anything but RSpec. The whole idea of acceptance testing, and the tools that are built around it, exist to create this beneficial relationship. It only makes sense to optimize that aspect.
Besides, is writing more step definitions really so painful to a developer? Is it more painful than having a client who is unhappy or unable and unwilling to communicate?
What are we communicating?
The other problem with the first example is that it doesn’t really communicate to either contingent, especially if they are coming into the project cold. “Sales Rep” is an implication. As such, there is another mental step to take there. What does the role “Sales Rep” represent? Oh yeah, a role that provokes a question on login. So just say that.
The tendency to write steps in the way of the first example is a result of trying test too much through the UI. Mitch gave the example of a signup form with many required fields, each doing a validation that will prevent the signup and display some kind of message. E.g. “Name is invalid”, “Email is invalid”, “Username is invalid”. The tendency is to write a step like:
I fill in "email" with "invalid_email_address"
And then to reuse this step for each required field. However, this proliferation of scenarios is going to simultaneously bore the consumer of the feature, and slow down your test suite. The correct place to test these validations is in RSpec unit tests. In Cucumber we should write:
I fill in the form with invalid data
Test that. Validate that a message appears, sign in is prohibited, and move on!
Another way to quickly bore and confuse Gherkin consumers? Insert unnecessary incidental details into a step definition. If we are testing authentication we might write things like:
When I log in as "firstname.lastname@example.org" with password "l0ck3d!"
Because that’s the data the login form accepts. But does it matter? Compare to this:
When I log in as "email@example.com"
Clarity is important. See how many incidental details can be removed, while retaining the scenario’s meaning. Is it important who we are? Are we testing anything about Dave specifically? How about this:
When I log in
Every effort should be made to pare down steps. When in doubt go with plain English. Get the scenario itself down the the magic number of three steps, i.e. “Given”, “When”, “Then”. Even the “Given” step can be extraneous. Eliminate the incidental!
Where should the programming be done?
Michael likened a step definition to an programming interface. We only need to have an idea of what the step does, we don’t care about the implementation (how it is done). But, because steps can act like functions that pull and apply parameters, the trap is often to move logic into them, and to use them like methods.
Cucumber is not a programming language. At every point the tendency should be to move logic and variables into the implementation and out of the steps themselves. Even the first scenario example introduces coupling between steps, it just does it on the Gherkin side instead of the Ruby side. Even though Cucumber supports it, do not call steps from steps in the implementation, call Ruby methods from steps.
In Cucumber, a developer should resist the temptation to always reuse. Don’t always sift through the step definition files to locate the existing step that already does the thing you are trying to do. One should write the step with the language that makes sense for the scenario.
On the other hand don’t fall into this trap:
When I sign in When I login When I authenticate
This is bad. You still need to work towards ubiquitous language. There is no increase of communication or context that rationalizes three different steps for the same login action.
It’s a balance thing. Remember, programming concepts are not our first priority here. The first priority with Cucumber is communication. Sometimes there are better ways to communicate the same functionality, depending on the context. In these cases, it’s best to avoid shoehorning the language of an existing step into a feature, just for the sake of re-use.
Is there a good way to organize step definitions?
Because you’re dealing with a big mess of globally applicable step definitions, it can be painful to locate any one implementation. Grepping seemed to be my goto method for locating step definitions, but I just verified that ctags in vim does the trick quite nicely. Still, it’s never clear exactly how to organize steps and features, or where to find them when you come onto a project.
If there’s a Gaslight convention, it’s to organize step definitions by domain object or concept, and features by user role. In this way, it’s easier to answer the question “what are the things this user can do”?
Our favorite new pattern here at Gaslight is the Page Object. Page Objects are a way to in step implementations to encapsulate step code into classes that provide an interface for creating test objects and interacting with the UI. The results are vastly simplified step definitions and better code organization.
Page Objects are probably a topic that warrant their own blog post. Until then I invite you to read up on them here.
Working with Cucumber and writing Gherkin are skills unto themselves, with best practices that can seem antithetical to those that define good software engineering. That is because the real goal of these tools is not testing. Cucumber and Gherkin are about communication and stakeholder engagement, and should be implemented with those goals in mind, first and foremost.