The primary goal of our CI system is to ensure that the master branch of rust-lang/rust is always in a valid state by passing our test suite.
From a high-level point of view, when you open a pull request at rust-lang/rust, the following will happen:
master branch.If you want to modify what gets executed on CI, see Modifying CI jobs.
Our CI is primarily executed on GitHub Actions, with a single workflow defined in .github/workflows/ci.yml, which contains a bunch of steps that are unified for all CI jobs that we execute. When a commit is pushed to a corresponding branch or a PR, the workflow executes the src/ci/citool crate, which dynamically generates the specific CI jobs that should be executed. This script uses the jobs.yml file as an input, which contains a declarative configuration of all our CI jobs.
Almost all build steps shell out to separate scripts. This keeps the CI fairly platform independent (i.e., we are not overly reliant on GitHub Actions). GitHub Actions is only relied on for bootstrapping the CI process and for orchestrating the scripts that drive the process.
In essence, all CI jobs run ./x test, ./x dist or some other command with different configurations, across various operating systems, targets, and platforms. There are two broad categories of jobs that are executed, dist and non-dist jobs.
rust-lang-ci2 S3 bucket and are available to be locally installed with the rustup-toolchain-install-master tool. The same builds are also used for actual releases: our release process basically consists of copying those artifacts from rust-lang-ci2 to the production endpoint and signing them.Based on an input event (usually a push to a branch), we execute one of three kinds of builds (sets of jobs).
After each push to a pull request, a set of pr jobs are executed. Currently, these execute the x86_64-gnu-llvm-X, x86_64-gnu-tools, pr-check-1, pr-check-2 and tidy jobs, all running on Linux. These execute a relatively short (~40 minutes) and lightweight test suite that should catch common issues. More specifically, they run a set of lints, they try to perform a cross-compile check build to Windows mingw (without producing any artifacts), and they test the compiler using a system version of LLVM. Unfortunately, it would take too many resources to run the full test suite for each commit on every PR.
Note on doc comments
Note that PR CI as of Oct 2024 by default does not try to run
./x doc xxx. This means that if you have any broken intradoc links that would lead to./x doc xxxfailing, it will happen very late into the full merge queue CI pipeline.Thus, it is a good idea to run
./x doc xxxlocally for any doc comment changes to help catch these early.
PR jobs are defined in the pr section of jobs.yml. Their results can be observed directly on the PR, in the “CI checks” section at the bottom of the PR page.
Before a commit can be merged into the master branch, it needs to pass our complete test suite. We call this an auto build. This build runs tens of CI jobs that exercise various tests across operating systems and targets. The full test suite is quite slow; it can take several hours until all the auto CI jobs finish.
Most platforms only run the build steps, some run a restricted set of tests; only a subset run the full suite of tests (see Rust's platform tiers).
Auto jobs are defined in the auto section of jobs.yml. They are executed on the auto branch under the rust-lang/rust repository, and the final result will be reported via a comment made by bors on the corresponding PR. The live results can be seen on the GitHub Actions workflows page.
At any given time, at most a single auto build is being executed. Find out more in Merging PRs serially with bors.
Sometimes we want to run a subset of the test suite on CI for a given PR, or build a set of compiler artifacts from that PR, without attempting to merge it. We call this a “try build”. A try build is started after a user with the proper permissions posts a PR comment with the @bors try command.
There are several use-cases for try builds:
@bors try @rust-timer queue command combination.By default, if you send a comment with @bors try, the jobs defined in the try section of jobs.yml will be executed. We call this mode a “fast try build”. Such a try build will not execute any tests, and it will allow compilation warnings. It is useful when you want to get an optimized toolchain as fast as possible, for a Crater run or performance benchmarks, even if it might not be working fully correctly. If you want to do a full build for the default try job, specify its job name in a job pattern (explained below).
If you want to run custom CI jobs in a try build and make sure that they pass all tests and do not produce any compilation warnings, you can select CI jobs to be executed by specifying a job pattern, which can be used in one of two ways:
try-job: <job pattern> directives to the PR description (described below) and then simply run @bors try. CI will read these directives and run the jobs that you have specified. This is useful if you want to rerun the same set of try jobs multiple times, after incrementally modifying a PR.jobs parameter of the try command: @bors try jobs=<job pattern>. This is useful for one-off try builds with specific jobs. Note that the jobs parameter has a higher priority than the PR description directives.@bors try jobs=job1,job2,job3.Each job pattern can either be an exact name of a job or a glob pattern that matches multiple jobs, for example *msvc* or *-alt. You can start at most 20 jobs in a single try build. When using glob patterns in the PR description, you can optionally wrap them in backticks (`) to avoid GitHub rendering the pattern as Markdown if it contains e.g. an asterisk. Note that this escaping will not work when using the @bors jobs= parameter.
The job pattern needs to match one or more jobs defined in the auto or optional sections of jobs.yml:
auto jobs are executed before a commit is merged into the master branch.optional jobs are executed only when explicitly requested via a try build. They are typically used for tier 2 and tier 3 targets.Using
try-jobPR description directives
Identify which set of try-jobs you would like to exercise. You can find the name of the CI jobs in
jobs.yml.Amend PR description to include a set of patterns (usually at the end of the PR description), for example:
This PR fixes #123456. try-job: x86_64-msvc try-job: test-various try-job: `*-alt`Each
try-jobpattern must be on its own line.Run the prescribed try jobs with
@bors try. As aforementioned, this requires the user to either (1) havetrypermissions or (2) be delegated withtrypermissions by@bors delegateby someone who hastrypermissions.Note that this is usually easier to do than manually edit
jobs.yml. However, it can be less flexible because you cannot adjust the set of tests that are exercised this way.
Try builds are executed on the try branch under the rust-lang/rust repository and their results can be seen on the GitHub Actions workflows page, although usually you will be notified of the result by a comment made by bors on the corresponding PR.
Multiple try builds can execute concurrently across different PRs, but there can be at most a single try build running on a single PR at any given time.
Note that try builds are handled using the new bors implementation.
If you want to modify what gets executed on our CI, you can simply modify the pr, auto or try sections of the jobs.yml file.
You can also modify what gets executed temporarily, for example to test a particular platform or configuration that is challenging to test locally (for example, if a Windows build fails, but you don‘t have access to a Windows machine). Don’t hesitate to use CI resources in such situations.
You can perform an arbitrary CI job in two ways:
pr section of jobs.yml to specify which CI jobs should be executed after each push to your PR. This might be faster than repeatedly starting try builds.To modify the jobs executed after each push to a PR, you can simply copy one of the job definitions from the auto section to the pr section. For example, the x86_64-msvc job is responsible for running the 64-bit MSVC tests. You can copy it to the pr section to cause it to be executed after a commit is pushed to your PR, like this:
pr: ... - image: x86_64-gnu-tools <<: *job-linux-16c # this item was copied from the `auto` section # vvvvvvvvvvvvvvvvvv - image: x86_64-msvc env: RUST_CONFIGURE_ARGS: --build=x86_64-pc-windows-msvc --enable-profiler SCRIPT: make ci-msvc <<: *job-windows-8c
Then you can commit the file and push it to your PR branch on GitHub. GitHub Actions should then execute this CI job after each push to your PR.
After you have finished your experiments, don't forget to remove any changes you have made to jobs.yml, if they were supposed to be temporary!
A good practice is to prefix [WIP] in PR title while still running try jobs and [DO NOT MERGE] in the commit that modifies the CI jobs for testing purposes.
Although you are welcome to use CI, just be conscious that this is a shared resource with limited concurrency. Try not to enable too many jobs at once; one or two should be sufficient in most cases.
CI services usually test the last commit of a branch merged with the last commit in master, and while that’s great to check if the feature works in isolation, it doesn’t provide any guarantee the code is going to work once it’s merged. Breakages like these usually happen when another, incompatible PR is merged after the build happened.
To ensure a master branch that works all the time, we forbid manual merges. Instead, all PRs have to be approved through our bot, bors (the software behind it is called homu). All the approved PRs are put in a merge queue (sorted by priority and creation date) and are automatically tested one at the time. If all the builders are green, the PR is merged, otherwise the failure is recorded and the PR will have to be re-approved again.
Bors doesn’t interact with CI services directly, but it works by pushing the merge commit it wants to test to specific branches (like auto or try), which are configured to execute CI checks. Bors then detects the outcome of the build by listening for either Commit Statuses or Check Runs. Since the merge commit is based on the latest master and only one can be tested at the same time, when the results are green, master is fast-forwarded to that merge commit.
Unfortunately, testing a single PR at a time, combined with our long CI (~2 hours for a full run), means we can’t merge a lot of PRs in a single day, and a single failure greatly impacts our throughput. The maximum number of PRs we can merge in a day is around ~10.
The long CI run times, and requirement for a large builder pool, is largely due to the fact that full release artifacts are built in the dist- builders. This is worth it because these release artifacts:
Some PRs don’t need the full test suite to be executed: trivial changes like typo fixes or README improvements shouldn’t break the build, and testing every single one of them for 2+ hours would be wasteful. To solve this, we regularly create a “rollup”, a PR where we merge several pending trivial PRs so they can be tested together. Rollups are created manually by a team member using the “create a rollup” button on the merge queue. The team member uses their judgment to decide if a PR is risky or not.
All CI jobs, except those on macOS and Windows, are executed inside that platform’s custom Docker container. This has a lot of advantages for us:
The build environment is consistent regardless of the changes of the underlying image (switching from the trusty image to xenial was painless for us).
We can use ancient build environments to ensure maximum binary compatibility, for example using older CentOS releases on our Linux builders.
We can avoid reinstalling tools (like QEMU or the Android emulator) every time, thanks to Docker image caching.
Users can run the same tests in the same environment locally by just running this command:
cargo run --manifest-path src/ci/citool/Cargo.toml run-local <job-name>
This is helpful for debugging failures. Note that there are only Linux Docker images available locally due to licensing and other restrictions.
The Docker images prefixed with dist- are used for building artifacts while those without that prefix run tests and checks.
We also run tests for less common architectures (mainly Tier 2 and Tier 3 platforms) in CI. Since those platforms are not x86, we either run everything inside QEMU, or we just cross-compile if we don’t want to run the tests for that platform.
These builders are running on a special pool of builders set up and maintained for us by GitHub.
Our CI workflow uses various caching mechanisms, mainly for two things:
The Docker images we use to run most of the Linux-based builders take a long time to fully build. To speed up the build, we cache them using Docker registry caching, with the intermediate artifacts being stored on ghcr.io. We also push the built Docker images to ghcr, so that they can be reused by other tools (rustup) or by developers running the Docker build locally (to speed up their build).
Since we test multiple, diverged branches (master, beta and stable), we can’t rely on a single cache for the images, otherwise builds on a branch would override the cache for the others. Instead, we store the images under different tags, identifying them with a custom hash made from the contents of all the Dockerfiles and related scripts.
The CI calculates a hash key, so that the cache of a Docker image is invalidated if one of the following changes:
We build some C/C++ stuff in various CI jobs, and we rely on Sccache to cache the intermediate LLVM artifacts. Sccache is a distributed ccache developed by Mozilla, which can use an object storage bucket as the storage backend.
With Sccache there's no need to calculate the hash key ourselves. Sccache invalidates the cache automatically when it detects changes to relevant inputs, such as the source code, the version of the compiler, and important environment variables. So we just pass the Sccache wrapper on top of Cargo and Sccache does the rest.
We store the persistent artifacts on the S3 bucket, rust-lang-ci-sccache2. So when the CI runs, if Sccache sees that LLVM is being compiled with the same C/C++ compiler and the LLVM source code is the same, Sccache retrieves the individual compiled translation units from S3.
During the years, we developed some custom tooling to improve our CI experience.
The build logs for rust-lang/rust are huge, and it’s not practical to find what caused the build to fail by looking at the logs. We therefore developed a bot called Rust Log Analyzer (RLA) that receives the build logs on failure, and extracts the error message automatically, posting it on the PR thread.
The bot is not hardcoded to look for error strings, but was trained with a bunch of build failures to recognize which lines are common between builds and which are not. While the generated snippets can be weird sometimes, the bot is pretty good at identifying the relevant lines, even if it’s an error we've never seen before.
The rust-lang/rust repo doesn’t only test the compiler on its CI, but also a variety of tools and documentation. Some documentation is pulled in via git submodules. If we blocked merging rustc PRs on the documentation being fixed, we would be stuck in a chicken-and-egg problem, because the documentation's CI would not pass since updating it would need the not-yet-merged version of rustc to test against (and we usually require CI to be passing).
To avoid the problem, submodules are allowed to fail, and their status is recorded in rust-toolstate. When a submodule breaks, a bot automatically pings the maintainers so they know about the breakage, and it records the failure on the toolstate repository. The release process will then ignore broken tools on nightly, removing them from the shipped nightlies.
While tool failures are allowed most of the time, they’re automatically forbidden a week before a release: we don’t care if tools are broken on nightly but they must work on beta and stable, so they also need to work on nightly a few days before we promote nightly to beta.
More information is available in the toolstate documentation.
To monitor the Rust CI, you can have a look at the public dashboard maintained by the infra team.
These are some useful panels from the dashboard:
To learn more about the dashboard, see the Datadog CI docs.
If you want to determine which bootstrap.toml settings are used in CI for a particular job, it is probably easiest to just look at the build log. To do this:
build.configure-args