Bazel Continuous Integration
Bazel uses Buildkite for continuous integration. The user interface and the orchestration of CI builds is fully managed by Buildkite, but Bazel brings its own CI machines. The buildkite folder contains all the scripts and configuration files necessary to setup Bazel's CI on Buildkite.
Bazel on Buildkite 101
When you first log into Buildkite you are presented with a list of pipelines. A pipeline is a template of steps that are executed either in sequence or in parallel and that all need to succeed in order for the pipeline to succeed. The Bazel organisation has dozens of pipelines. Here are a selected few:
- The bazel postsubmit pipeline builds and tests each commit to Bazel's repository on all supported platforms.
- The bazel presubmit pipeline is triggered on every pull request to Bazel.
- The rules_go postsubmit pipeline is triggered on every commit to the rules_go repository.
- The TensorFlow pipeline builds and tests TensorFlow at
HEADevery four hours.
When you click on a pipeline you can see the last few builds of this pipeline. Clicking on a build then gives you access to the details of the build. For example, the below image shows a failed build step on Ubuntu 16.04.
One can see which tests failed by clicking on the Test section. In the below example, the
//src/test/shell/bazel:external_path_test was flaky as it failed in 1 out of 5 runs.
You can view the failed test attempt's
test.log file in the Artifacts tab.
Bazel accepts contributions via pull requests. Contributions by members of the bazelbuild organisation as well as members of individual repositories (i.e. rule maintainers) are whitelisted automatically and will immediately be built and tested on Buildkite.
An external contribution, however, first needs to be verified by a project member and therefore will display a pending status named Verify Pull Request.
A member can verify a pull request by clicking on Details, followed by Verify Pull Request.
Please vet external contributions carefully as they can execute arbitrary code on our CI machines
Build and Test Results
After a pull request has been built and tested, the results will be displayed as a status message on the pull request. A detailed view is available when clicking on the corresponding Details link. Click here for an example.
Bazel downstream projects is red? Use culprit finder to find out which bazel commit broke it!
First you should check if the project is green with the latest Bazel release. If not, probably it's their commits that broke the CI.
If a project is green with release Bazel but red with Bazel nightly, it means some Bazel commit broke it, then culprit finder can help!
Create "New Build" in the Culprit Finder project with the following environment variable:
- PROJECT_NAME (The project name must exists in DOWNSTREAM_PROJECTS in bazelci.py)
- PLATFORM_NAME (The platform name must exists in PLATFORMS in bazelci.py)
- GOOD_BAZEL_COMMIT (A full Bazel commit, Bazel built at this commit still works for this project)
- BAD_BAZEL_COMMIT (A full Bazel commit, Bazel built at this commit fails with this project)
- (Optional) NEEDS_CLEAN (Set NEEDS_CLEAN to
bazel clean --expungebefore each build, this will help reduce flakiness)
PROJECT_NAME=rules_go PLATFORM_NAME=ubuntu1404 GOOD_BAZEL_COMMIT=b6ea3b6caa7f379778e74da33d1bd0ff6477f963 BAD_BAZEL_COMMIT=91eb3d207714af0ab1e5812252a0f10f40d6e4a8
Note: Bazel commit can only be set to commits after 63453bdbc6b05bd201375ee9e25b35010ae88aab, Culprit Finder needs to download Bazel at specific commit, but we didn't prebuilt Bazel binaries before this commit.
Configuring a Pipeline
Each pipeline is configured via a Yaml file. This file either lives in
$PROJECT_DIR/.bazelci/presubmit.yml (for presubmits) or in an arbitrary location whose path or URL is passed to the CI script (as configured in the Buildkite settings of the respective pipeline). Projects should store the postsubmit configuration in their own repository, but we keep some configurations for downstream projects in https://github.com/bazelbuild/continuous-integration/tree/master/buildkite/pipelines.
The most important piece of the configuration file is the
tasks dictionary. Each task has a unique key, a platform and usually some build and/or test targets:
--- tasks: ubuntu_build_only: platform: ubuntu1604 build_targets: - "..." windows: platform: windows build_targets: - "..." test_targets: - "..."
If there is exactly one task per platform, you can omit the
platform field and use its value as task ID instead. The following code snippet is equivalent to the previous one:
--- tasks: ubuntu1604: build_targets: - "..." windows: build_targets: - "..." test_targets: - "..."
There are more advanced mechanisms for configuring individual tasks:
- The CI script can run arbitrary commands before building or testing.
Simply add the
shell_commandsfield (all other platforms).
run_targetsfield contains a list of Bazel targets that should be run before building.
test_flagsfields contain lists of flags that should be used when building or testing (respectively).
- You can specify a display name in the
namefield. The resulting job will display this value and an emoji that represents the platform.
--- tasks: ubuntu1404: shell_commands: - rm -f obsolete_file run_targets: - "//whatever" build_flags: - "--define=ij_product=clion-latest" build_targets: - "..." test_flags: - "--define=ij_product=clion-latest" test_targets: - ":clwb_tests" windows: name: "some :emoji:" batch_commands: - powershell -Command "..." build_targets: - "..."
Most existing configuration use the legacy format with a "platforms" dictionary:
--- platforms: ubuntu1404: build_targets: - "..." test_targets: - "..."
The new format expects a "tasks" dictionary instead:
--- tasks: arbitrary_id: platform: ubuntu1404 build_targets: - "..." test_targets: - "..."
In this case we can omit the
platforms field since there is a 1:1 mapping between tasks and platforms. Consequently, the format looks almost identical to the old one:
--- tasks: ubuntu1404: build_targets: - "..." test_targets: - "..."
The CI script still supports the legacy format, too.
Using a specific version of Bazel
The CI uses Bazelisk to support older versions of Bazel, too. You can specify a Bazel version for each pipeline (or even for individual platforms) in the pipeline Yaml configuration:
--- bazel: 0.20.0 platforms: windows: build_targets: - "..." macos: build_targets: - "..." ubuntu1404: bazel: 0.18.0 build_targets: - "..." [...]
In this example the jobs on Windows and MacOS would use 0.20.0, whereas the job on Ubuntu would run 0.18.0. Please see the Bazelisk documentation for a list of all supported version values.
Running Buildifier on CI
For each pipeline you can enable Buildifier to check whether all BUILD, BUILD.bazel and .bzl files comply with the standard formatting convention. Simply add the following code to the top of the particular pipeline configuration:
--- buildifier: true [...]
As a consequence, every future build for this pipeline will contain an additional "Buildifier" step that runs the latest version of Buildifier in "lint" mode.