For the last two weeks I've been experimenting with Behavior-Driven Development (BDD) in order to find out what it takes for the Quality Assurance department to switch from using the Waterfall method to BDD. Here are my initial observations and thoughts for further investigation.
Developing an entire Linux distribution (or any large product for that matter) is a very complicated task. Traditionally QA has been involved in writing the test plans for the proposed technology updates, then execute and maintain them during the entire product life-cycle reporting and verifying tons of bugs along the way. From the point of view of the entire product the process is very close to the traditional waterfall development method. I will be using the term waterfall to describe the old way of doing things and BDD the new one. In particular I'm referring to the process of analyzing the proposed feature set for the next major version of the product (e.g. Fedora) and designing the necessary test plans documents and test cases.
To get an idea about where does QA join the process see the Fedora 24 Change set. When the planning phase starts we are given these "feature pages" from which QA needs to distill test plans and test cases. The challenges with the waterfall model are that QA joins the planning process rather late and there is not enough time to iron out all the necessary details. Add to this the fact that feature pages are often incomplete and vaguely described and sometimes looking for the right answers is the hardest part of the job.
QA and BDD
Right now I'm focusing on using the Gherkin Given-When-Then language to prepare feature descriptions and test scenarios from the above feature pages. You can follow my work on GitHub and I will be using them as examples below. Also see examples from my co-workers 1, 2.
With this experiment I want to verify how hard/easy it is for QA to write test cases using BDD style documents and how is that different from the traditional method. Since I don't have any experience (nor bias) towards BDD I'm documenting my notes and items of interest.
It took me about 2 hours to get started. The essence of Gherkin is the Action & Response mechanism. Given the system under test (SUT) is in a known state and when an action is taken then we expect something to happen in response to the action. This syntax made me think from the point of view of the user. This way it was very easy to identify different user roles and actions which will be attempted with the SUT. This also made my test scenarios more explicit compared to what is described in the wiki pages. IMO being explicit when designing tests is a good thing. I like it that way.
OTOH the same explicitness can be achieved with the waterfall method as well. The trouble is that this is often overlooked because we're not in the mindset to analyze the various user roles and scenarios. When writing test cases with waterfall the mindset is more focused on the technical features, e.g. how the SUT exactly works and we end up missing important interactions between the user and the system. At least I can recall a few times that I've made that mistake.
Tagging the scenarios is a good way of indicating which scenario covers which roles. Depending on the tools you use it should be possible to execute test scenarios for different roles (tags). In waterfall we need to have a separate test plan for each user role, possibly duplicating some of the test cases across test plans. A bit redundant but more importantly easier to forget the bigger picture.
Big, Small & Undefined
BDD originates from TDD which in turn relies heavily on unit testing and automation. This makes it very easy to use BDD test development (and even automate) for self contained changes, especially ones which affect only a single component (e.g. a single program). From the Fedora 24 changes such are for example the systemd and system-python split.
I happen to work in a team where we deal with large changes, which affect multiple components and infrastructure. Both the Pungi Refactor and Layered Docker Image Build Service for which I've written BDD style test scenarios are of this nature. This leads to the following issues:
- QA doesn't always have the entire infrastructure stack in a staging environment for testing so we need to test on the live infra;
- QA doesn't always have the necessary access permissions to execute the tests and in some cases never will. For example it is very unlikely that QA will be able to build a test release and push that for syncing to the mirrors infrastructure to verify that there are no files left behind;
- Not being able to test independently means QA has to wait for something to happen then verify the results (e.g. rel-eng builds new Docker images and pushes them live). When something breaks this testing is often too late.
Complex changes are often not described into detail. As they affect multiple infrastructure layers and components sometimes it is not known what the required changes need to be. That's why we implement them in stages and have contingency plans. However this makes it harder for QA to write the tests. Btw this is the same regardless of which development method is used. The good thing is that by forcing you to think from the POV of the user and in terms of action & response BDD helps identify these missing bits faster.
For example, with Pungi (Fedora distro build tool), the feature pages says that the produced directory structure will be different from previous releases but it doesn't say what is going to be different so we can't really test that. I know from experience that this may break tools which rely on this structure like virt-manager and anaconda and have added simple sanity tests for them.
In the Docker feature page we have functional requirement for automatic image rebuilds if one of the underlying components (e.g. RPM package) changes. This is not described in details and so is the test scenario. I can easily write a separate BDD feature document for this functionality alone.
With the waterfall model when a feature isn't well defined QA often waits for the devel team to implement the basic features and then writes test cases based on the existing behavior. This is only good for regression testing the next version but it can't show you something that is missing because we're never going to look for it. BDD makes it easier to spot when we need better definitions of scope and roles, even better functional requirements.
Automation and Integration
Having a small SUT is nice. For example we can easily write a test script to install, upgrade and query RPM packages and verify the systemd package split. We can easily prepare a test system and execute the scripts to verify the expected results.
OTOH complex features are hard to integrate with BDD automation tools. For the Docker Image Build Service the straight forward script would be to start building a new image, then change an underlying component and see if it gets rebuilt, then ensure all the content comes from the existing RPM repos, then push the image to the Docker registry and verify it can be used by the user, etc, etc. All of these steps take a non-trivial amount of time. Sometimes hours. You can also execute them in parallel to save time but how do you sync back the results ?
My preference for the moment is to kick-off individual test suites for a particular BDD scenario and then aggregate the results back. This also has a side benefit - for complex changes we can have layered BDD feature documents, each one referencing another feature document. Repeat this over and over until we get down to purely technical scenarios which can be tested easily. Once the result are in go back the chain and fill-in the rest. This way we can traverse all testing activities from the unit testing level up to the infrastructure level.
I actually like the back & forth traversing idea very much. I've always wanted to know how does each individual testing effort relate to the general product development strategy and in which areas the product is doing well or not. You can construct the same chain of events with waterfall as well. IMO BDD just makes it a bit more easier to think about it.
Another problem I faced is how do I mark the scenarios as out of scope for the current release ? I can tag them or split them into separate files or maybe something else? I don't know which one is the best practice. In waterfall I'll just disable the test cases or move them into a separate test plan.
I will be writing more BDD test definitions in the upcoming 2 weeks to get more experience with them. I still don't have a clear idea how to approach BDD test writing when given a particular feature to work on. So far I've used the functional requirements and items of concern (when present), in the feature pages, as a starting point for my BDD test scenarios.
I also want to get more feedback from the development teams and product management folks.
BDD style test writing puts the tester into a mind set where it is easier to see the big picture by visualizing different user roles and scenarios. It makes it easier to define explicit test cases and highlights missing details. It is easier for QA to join early in the planning process by defining roles and thinking about all the possible interactions with the SUT. This is the biggest benefit for me!
Self-contained changes are easier to describe and test automatically.
Bigger and complex features are harder to describe and even harder to automate in one piece. Divide and conqueror is our best friend here!