Raising the Continuous Integration Bar


So we implemented Continuous Integration. It gives us the confidence to deliver – great! Though, we are in a suboptimal situation. We do not yet have all the benefits that Continuous Integration should provide. How can we get the most out of it? Which additional practices are we missing and should we acquire? How can we raise the bar?


Raising the Continuous Integration Bar

Push Every Day is one of the six foundational practices to get started with Continuous Integration. No matter how, if we want to leverage the benefits of Continuous Integration to the fullest, integrating once per day will not make it. That is pertinent but inadequate. To raise the bar to the next level, we have to Commit Frequently. After all, the central premise of Continuous Integration is integrating early and often on Mainline.

When we Commit Frequently we reduce the size of our changes. Merge conflicts are least likely and broken builds are infrequent.

It also reduces risks. If the build happens to break, it is fairly easy to find the root cause because the changeset is so small. Probably also, because we only introduced the failing change just minutes ago. We still have enough context to fix the build readily. Getting the build back to green within ten minutes becomes attainable.

If we have to revert a failing commit, that too, will be effortless. Because the change is so small, we are not plagued by the sunk cost fallacy. Thus, deciding to revert will be straightforward.

Consequently, we should commit at least once an hour, preferably multiple times per hour. Thoughtful engineers commit and push after every tiny refactoring. This has the benefit of communicating the refactoring in real-time with the team, avoiding surprises for the team when releasing a series of refactorings that evolve into a complete redesign.

Once we commit and push every so often into the remote Mainline, multiple times per hour, it becomes apparent that version control branches, and specifically Feature Branches with Pull Requests, becomes untenable. The cognitive overhead is way out of balance compared to the illusive benefits. Additionally, Pull Requests will bring all the benefits we built up with the practices that bring Continuous Integration to a blatant halt.

Naturally, Commit Frequently creates a gentle design pressure to work in much more, many smaller steps, to keep the code base even more decoupled, to further speed up the build, and lastly to increasingly hide unfinished functionality. This creates a virtuous circle resulting in committing all the more frequently.

To Make all Changes in Small Increments it is vital to Decouple the Code Base. When a codebase is too coupled, it becomes tough to adopt incremental software engineering skills. Any change will rip apart the application and prevent the application from working all the time. We find ourselves not releasing anymore at any time, incurring delays in delivery and an increased opportunity cost.

Adopting Ports and Adapters together with Simple Design and intentional code duplication (through Expand-Contract) helps to decouple the code base. Interestingly, these patterns not only amplify quality and improve maintainability, they also optimise the required engineering time for new functionality. It becomes cheaper to introduce changes.

Inherently, when a code base is decoupled, we can Commit More Frequently.

Most often, a feature takes a fair amount of time to implement. It involves a series of commits. We grow the feature small step by small step, commit by commit. How can we ensure the unfinished functionality is not released to the end user?

The classic approach is to use Feature Branches. As long as the feature is not done, the branch is not merged into Mainline. Intrinsically, the feature is hidden using the version control system as a manual toggling mechanism. This has a downside. The feature is not integrated with all the other ongoing changes while the feature is being implemented. We are blind to whether the feature will cause any integration problems until the feature is finished and merged into Mainline. This delays feedback. Inevitably, it negatively impacts quality and necessarily increases the lead time to market.

Instead, we could plainly Hide Unfinished Functionality. This is likely the most uncomplicated practice to adopt. It is entirely acceptable to have unfinished functionality sitting in production as long as it is not discoverable and does not introduce any security risks. Commonly, we do not need elaborate Feature Toggles for that. For instance, when adding a new screen to an existing user interface we simply do not wire the screen into the navigation. We only do that at the end, once the screen is ready. The same is true for new API endpoints or backend services. As long as they are not ready, we do not document them or we do not use the backend service.

But, in all honesty, Feature Toggles will be necessary when changing existing screens or modifying the behaviour of existing backends. Managing toggles requires mindfulness and consciousness if we do not want to self-sabotage ourselves.

In some cases, we will have to perform a large refactoring that can rip apart the application for an extended period. Let us say, we want to replace a library or a framework with another or we need to break backwards compatibility on a certain service. Here too, the classic approach is to use Feature Branches. Again, the problem is we are not integrating. In the meantime, while the large refactoring is happening, new functionality is added. When the refactoring is finally ready, the integration will be tedious and time-consuming. Accordingly, delivery comes to a halt and delivery lead times go through the roof. On top of that, we have created a massive amount of inventory that we do not dare to integrate as we fear how complicated that will be.

How can we do this more smoothly, in a more effective way, without impacting the flow of delivery? Adopt Expand-Contract. This is the intentional code duplication we mentioned earlier. Once we understand Expand-Contract, it is gold!

Truthfully, Expand-Contract will introduce a certain level of complexity. We will have to think harder. We might move slower. But, it brings the tremendous advantage of delivering new functionality while a large-scale refactoring is happening that takes days, weeks or even months to complete. The application keeps working at all times, allowing us to perform on-demand releases anytime. At no single moment, we are blocked. The flow of delivery steadily continues.

To Commit Frequently into Mainline, we must Run the Local Build repeatedly before each commit to satisfy Agree as a Team to Never Break the Build. Additionally, when engineers commit multiple times per hour, they execute the Local Build multiple times per hour. That means we ought to Have a Fast Build. This is primordial! But also hard work. Lots of teams struggle with that. Often teams hide slow builds behind version control branches and remote Pull Request builds that in turn introduce a decent amount of context switching and fatigue.

When the build is too slow, two things can happen. Either teams do not tend to run the Local Build before committing. We then run the risk of introducing regressions, thus impacting delivery throughput. Or teams tend to execute the build less often. In that case, batch work is introduced. We know from Lean Manufacturing that it drives down quality and consequently inflates time to market.

But, what is fast? We focus on a Local Build (and thus also a Commit Build) of five minutes but no longer than ten minutes. If we can bring it to 30 seconds, it will avoid any hallway sword-fighting sessions.

xkcd: Compiling

To conclude, if we want to raise the bar, it is fundamental to Have a Fast Build. This is crucial as it allows us to commit all the more frequently, enabling us to work in increasingly smaller steps. But to Commit Frequently, also requires to Have a Decoupled Code Base and understanding we have to Hide Unfinished Functionality. Lastly, Adopt Expand-Contract helps us to refactor in small increments and to commit frequently when refactoring and delivering new functionality at the same time.

Definitions

Mainline

The Mainline is the line of development in Version Control which is the reference from which system builds are created that feed into a deployment pipeline.

For CVS and SubVersion, this is trunk. For Git, this is the remote main branch. For Mercurial, this is the remote default branch.

Commit Build

The Commit Build is a build performed during the first stage of the Deployment Pipeline or the central build server. It involves checking out the latest sources from Mainline and at minimum compiling the sources, running a set of Commit Tests, and building a binary artefact for deployment.

Commit Tests

The Commit Tests comprise all of the Unit Tests along with a small simple smoke test suite executed during the Commit Build. This smoke test suite includes a few simple Integration and Acceptance Tests deemed important enough to get early feedback.