This directory contains all of our automatically triggered workflows.
Our top level test_runner.yml is responsible for kicking off all tests, which are represented as reusable workflows. This is carefully constructed to satisfy the design laid out in go/protobuf-gha-protected-resources (see below), and duplicating it across every workflow file would be difficult to maintain. As an added bonus, we can manually dispatch our full test suite with a single button and monitor the progress of all of them simultaneously in GitHub's actions UI.
There are five ways our test suite can be triggered:
Post-submit tests (push): These are run over newly submitted code that we can assume has been thoroughly reviewed. There are no additional security concerns here and these jobs can be given highly privileged access to our internal resources and caches.
Pre-submit tests from a branch (push_request): These are run over every PR as changes are made. Since they are coming from branches in our repository, they have secret access by default and can also be given highly privileged access. However, we expect many of these events per change, and likely many from abandoned/exploratory changes. Given the much higher frequency, we restrict the ability to write to our more expensive caches.
Pre-submit tests from a fork (push_request_target): These are run over every PR from a forked repository as changes are made. These have much more restricted access, since they could be coming from anywhere. To protect our secret keys and our resources, tests will not run until a commit has been labeled safe to submit. Further commits will require further approvals to run our test suite. Once marked as safe, we will provide read-only access to our caches and Docker images, but will generally disallow any writes to shared resources.
Continuous tests (schedule): These are run on a fixed schedule. We currently have them set up to run daily, and can help identify non-hermetic issues in tests that don't get run often (such as due to test caching) or during slow periods like weekends and holidays. Similar to post-submit tests, these are run over submitted code and are highly privileged in the resources they can use.
Manual testing (workflow_dispatch): Our test runner can be triggered manually over any branch. This is treated similarly to pre-submit tests, which should be highly privileged because they can only be triggered by the protobuf team.
While Bazel handles code generation seamlessly, we do support build systems that don‘t. There are a handful of cases where we need to check in generated files that can become stale over time. In order to provide a good developer experience, we’ve implemented a system to make this more manageable.
Stale files should have a corresponding staleness_test Bazel target. This should be marked manual to avoid getting picked up in CI, but will fail if files become stale. It also provides a --fix flag to update the stale files.
Bazel tests will never depend on the checked-in versions, and will generate new ones on-the-fly during build.
Non-Bazel tests will always regenerate necessary files before starting. This is done using our bash and docker actions, which should be used for any non-Bazel tests. This way, no tests will fail due to stale files.
A post-submit job will immediately regenerate any stale files and commit them if they've changed.
A scheduled job will run late at night every day to make sure the post-submit is working as expected (that is, it will run all the staleness tests).
The regenerate_stale_files.sh script is the central script responsible for all the re-generation of stale files.
Because we need secret access to run our tests, we use the pull_request_target event for PRs coming from forked repositories. We do checkout the code from the PR‘s head, but the workflow files themselves are always fetched from the base branch (that is, the branch we’re merging to). Therefore, any changes to these files won't be tested, so we explicitly ban PRs that touch these files.
We have a number of different caching strategies to help speed up tests. These live either in GCP buckets or in our GitHub repository cache. The former has a lot of resources available and we don't have to worry as much about bloat. On the other hand, the GitHub repository cache is limited to 10GB, and will start pruning old caches when it exceeds that threshold. Therefore, we need to be very careful about the size and quantity of our caches in order to maximize the gains.
As described in https://bazel.build/remote/caching, remote caching allows us to offload a lot of our build steps to a remote server that holds a cache of previous builds. We use our GCP project for this storage, and configure every Bazel call to use it. This provides substantial performance improvements at minimal cost.
We do not allow forked PRs to upload updates to our Bazel caches, but they do use them. Every other event is given read/write access to the caches. Because Bazel behaves poorly under certain environment changes (such as toolchain, operating system), we try to use finely-grained caches. Each job should typically have its own cache to avoid cross-pollution.
When Bazel starts up, it downloads all the external dependencies for a given build and stores them in the repository cache. This cache is separate from the remote cache, and only exists locally. Because we have so many Bazel dependencies, this can be a source of frequent flakes due to network issues.
To avoid this, we keep a cached version of the repository cache in GitHub's action cache. Our full set of repository dependencies ends up being ~300MB, which is fairly expensive given our 10GB maximum. The most expensive ones seem to come from Java, which has some very large downstream dependencies.
Given the cost, we take a more conservative approach for this cache. Only push events will ever write to this cache, but all events can read from them. Additionally, we only store three caches for any given commit, one per platform. This means that multiple jobs are trying to update the same cache, leading to a race. GitHub rejects all but one of these updates, so we designed the system so that caches are only updated if they've actually changed. That way, over time (and multiple pushes) the repository caches will incrementally grow to encompass all of our dependencies. A scheduled job will run monthly to clear these caches to prevent unbounded growth as our dependencies evolve.
In order to speed up non-Bazel builds to be on par with Bazel, we make use of ccache. This intercepts all calls to the compiler, and caches the result. Subsequent calls with a cache-hit will very quickly short-circuit and return the already computed result. This has minimal affect on any single job, since we typically only run a single build. However, by caching the ccache results in GitHub's action cache we can substantially decrease the build time of subsequent runs.
One useful feature of ccache is that you can set a maximum cache size, and it will automatically prune older results to keep below that limit. On Linux and Mac cmake builds, we generally get 30MB caches and set a 100MB cache limit. On Windows, with debug symbol stripping we get ~70MB and set a 200MB cache limit.
Because CMake build tend to be our slowest, bottlenecking the entire CI process, we use a fairly expensive strategy with ccache. All events will cache their ccache directory, keyed by the commit and the branch. This means that each PR and each branch will write its own set of caches. When looking up which cache to use initially, each job will first look for a recent cache in its current branch. If it can't find one, it will accept a cache from the base branch (for example, PRs will initially use the latest cache from their target branch).
While the ccache caches quickly over-run our GitHub action cache, they also quickly become useless. Since GitHub prunes caches based on the time they were last used, this just means that we'll see quicker turnover.
Bazelisk will automatically download a pinned version of Bazel on first use. This can lead to flakes, and to avoid that we cache the result keyed on the Bazel version. Only push events will write to this cache, but it's unlikely to change very often.
Instead of downloading a fresh Docker image for every test run, we can save it as a tar and cache it using docker image save and later restore using docker image load. This can decrease download times and also reduce flakes. Note, Docker's load can actually be significantly slower than a pull in certain situations. Therefore, we should reserve this strategy for only Docker images that are causing noticeable flakes.
The actions/setup-python action we use for Python supports automated caching of pip dependencies. We enable this to avoid having to download these dependencies on every run, which can lead to flakes.
We've defined a number of custom actions to abstract out shared pieces of our workflows.
Bazel use this for running all Bazel tests. It can take either a single Bazel command or a more general bash command. In the latter case, it provides environment variables for running Bazel with all our standardized settings.
Bazel-Docker nearly identical to the Bazel action, this additionally runs everything in a specified Docker image.
Bash use this for running non-Bazel tests. It takes a bash command and runs it verbatim. It also handles the regeneration of stale files (which does use Bazel), which non-Bazel tests might depend on.
Docker nearly identical to the Bash action, this additionally runs everything in a specified Docker image.
ccache this sets up a ccache environment, and initializes some environment variables for standardized usage of ccache.
Cross-compile protoc this abstracts out the compilation of protoc using our cross-compilation infrastructure. It will set a PROTOC environment variable that gets automatically picked up by a lot of our infrastructure. This is most useful in conjunction with the Bash action with non-Bazel tests.