Recently, I’ve merged several AnySoftKeyboard repositories into a monorepo/monolithic-repository - I’m a believer in the monorepo process. During that process, I also refined the CI/CD process and wanted to give an overview of the process and the logic behind that.

This is the first part out of three:

  1. Continuous Integration, a.k.a CI.
  2. Continuous Deployment, a.k.a CD.
  3. Complimenting workflows.


The CI/CD flow in AnySoftKeyboard is as follows: Pull-request submitted -> Checks passed -> Code reviewed -> Pull-request merged -> Checks run on master -> Deployed to users.

The first five are the CI part, and the last is the CD part.

Why CI - Continuous Integration

Continuous Integration is pretty much an industry-standard, and well-accepted. In essence, for every commit (or a set of commits, e.g., pull-request), we run unit-tests, static-analysis tools, and code-style verifications. These checks will ensure three things:

  1. You are confident that your change did not break anything - hence you do not need to check every possible scenario and use-case.
  2. The next developer that picks up your changes gets a healthy, readable, and working code, feeling confident that they can start working on their code-change.
  3. With some confidence, any feature and use-case are still working for end-users.

You’ll notice that the keyword here is confidence. Continuous Integration gives us confidence that the HEAD commit of master is valid and working as intended.

Pull Requests

Every change will be proposed as a pull-request. The system will then perform the following:

  1. Assigning all relevant code-owners as reviewers according to the CODEOWNERS file. Assigning is done automatically by github.
  2. Running all checks. Defined in the checks workflow.
  3. Block merging until checks pass and code-review approved. This is configured in the protected branch setting in the repository settings page.


Checks are what makes us confident that our code works and does not break any existing functionality is still working as before this proposed code-change. Several types of checks run on each code-change.


The most common check that we run is unit-tests. Essentially, unit-tests will verify that distinct pieces of code are doing what they suppose to do correctly. For example, a test could verify that a Dictionary implementation can look up words and find typos.

We also run integration-tests, which ensure that different parts of the code-base work with each other as intended. For example, a test that given a user-interaction with the keyboard-view a specific text is printed in the Android’s text-box.

We write tests in two main cases:

  1. New functionality that we want to verify that is doing was we wanted: for each possible input category (or for each internal code branch), we have a test.
  2. A bug fix that we want to verify that won’t get regress back. These types of tests provide us with the most confidence in our code.

Note that tests cover the behavior of the code, and only the branches we individually coded in the test. So, of course, this is not complete.


Even if your code works under typical use-cases and all tests pass, it does not mean it is correct. Static-Analysis tools can search your code-base for common programming issues. Such issues could be an incorrect use of an external API or assuming something is not-null when it could.

In AnySoftKeyboard, we use:

  • Error-Prone, with an Android-specific configuration. Error-Prone is a fast checker that can find many Java-related code-smells. It is widely popular and actively maintained.
  • Lint with an AnySoftKeyboard-specific configuration. Lint is the gold-standard in Android-centric code and resources static-analysis. It is well-known and actively maintained.
  • checkstyle with an AnySoftKeyboard-specific configuration. Another well-known Java code-analysis. We use it to find illegal/undesirable code usage and files.
  • Google-Java-Format. This tool ensures that the code-style across the code-base is defined well and maintained. It also supports automatic formatting, so win-win.

Of course, many other tools can be used. A few notable mentions:

For us, we decided to focus on a few, very popular and fast, static-analysis to reduce the complexity of the checks-suite and the requirements from the developers.


A side-note. A lot is going on here; compiling production code, debug code and tests, running tests, and several tools on top of that code. This process could be slow, both in the CI environment and in the local developer environment. To that end, it is crucial to pick a build-system and tools that support caching (that is, do not execute actions if the input has not changed) and configure them correctly. Luckily, Gradle has an excellent caching mechanism (as long as your system is not too complicated): it will cache compilation steps and test-runs. Error-Prone is running inside the compilation step, so we get caching for free here. Lint also has an excellent caching mechanism, relying on Gradle’s caching mechanism.

Another benefit of caching is that faster feedback makes developers happier. Like, much much happier.

Persistent Cache

Gradle will store its cache on disk (there is an option to use remote-cache, under a specified path. Gradle is using this disk-cache in sequential runs to skip actions that have run in the past. This state is called warm-cache - skipping a step because you already have the result of it in the cache.

Keeping your local cache warm, is easy, and happens as you build locally. But, on CI, it’s a bit harder since every time you run a job on CI, you will get a different machine, one that does not have any cache related to your build. Run checks without cache means much slower checks.

Since we are using Github Workflows, we can specify where is our cache is located, how to identify its version, etc. See our checks-workflow configuration for details.

For AnySoftKeyboard, running the PR checks without cached can take up to 20 minutes, while about 8 minutes with the cache warmed.


Caching relies on the assumption that if the input is the same, the action will produce the same output. This assumption is true only if: The inputs are defined correctly. The action only uses the inputs to produce the output. If both these guidelines apply, then the action is defined as hermetic.

As a rule, Gradle is not a hermetic build-system (bazel, for example, is), but it does an outstanding job nonetheless.

To reduce the chances of a misbehaving action, we’ll use Docker in CI, so Gradle will have the same environment during its execution. Same Java, same bash, same SDK, same NDK, etc.

We are using an image that is built explicitly for AnySoftKeyboard. This image uses Java10 (this helps ErrorProne run faster), has the required SDK, NDK and build-tools all pre-installed and ready to go. Check out the Dockerfile for this image and the CI build script on GitHub.

As a side note, I would mention that it’s pretty easy to create a custom image and store it in on Docker Hub.

To be continued

What will happen to our pull-request? How will it be integrated into the master branch? And, how will it reach end-users? Keep watch for the next parts.

blog comments powered by Disqus