The Flutter Engine Build Definition Language describes a build on CI by defining a combination of sub-builds, archives, generators and dependencies. It makes it simple to shard sub-builds by mapping build inputs to workflows, and listing the sub-build-generated artifacts explicitly. The Build Definition Language, Engine Recipes V2 and the generation of artifacts using GN+Ninja set the groundwork for efficient builds with dependency reusability.
Author: Godofredo Contreras (godofredoc)
Go Link: flutter.dev/go/engine-build-definition-language
Created: 01/2023 / Last updated: 04/2023
- recipes - domain specific language for specifying sequences of subprocess calls in a cross-platform and testable way.
- Generator - scripts in Dart, python or bash that combine the output of sub-builds to generate artifacts.
- Builder - a combination of configuration, recipes and a given commit to build and test artifacts.
- Build - a builder running with specific properties, repository and commit.
- GN - a meta-build system that generates build files for Ninja.
- Ninja - Ninja is a small build system with a focus on speed.
- CAS - a service that stores arbitrary binary blobs addressed by (hash of) their content. It is specialized for low latency, high volume query/read/write operations.
Engine build definition files using the Build Definition Language can be found in the flutter/engine/ci/builders directory.
The engine orchestrator recipe reads each file in that directory, shards their builds, collects artifacts and uploads them to the Google Cloud Storage bucket.
The .ci.yaml file at the
root of the flutter/engine repository puts all the components together.
Builds are specified in that file using a property pointing to the build definition
file to be used by engine_v2 recipes. Full documentation of the .ci.yaml file format
can be found in the Cocoon repository here.
The following is a sample build configuration referencing
android_aot_engine.json
in the config_name under properties:
- name: Mac mac_android_aot_engine
recipe: engine_v2/engine_v2
timeout: 60
properties:
config_name: mac_android_aot_engine
$flutter/osx_sdk : >-
{ "sdk_version": "14e300c" }
To keep the build definition language simple the following assumptions were made during its design:
- A build can be expressed as a set of independent sub-builds.
- A sub-build can be defined as a sequence of a
gnconfiguration step, aninjabuild step, followed by self-contained test scripts, and self-contained generator scripts. All the elements are optional allowing to use gn+ninja without generators or generators without gn+ninja. - All the sub-builds required by a global generator are defined within the same configuration file.
The build configuration is a json file containing a list of builds, tests, generators and archives. The following is an example of an empty configuration file:
{
"builds": [],
"tests": [],
"generators": {
"tasks": []
},
"archives": [
]
}Note: tests, generators and archives can be omited if empty.
Build configuration files have to be checked into the engine_checkout/ci/builder directory where engine v2 recipes will be reading them from.
Configurations with a single build are supported. Single build configurations are located have to be checked into the engine_checkout/ci/builder/standalone
A configuration file defines a top-level builder that will show up as a column in the Flutter Dashboard.
A build is a dictionary with a gn command, a ninja command, zero or more generator commands, zero or more local tests, zero or more local generators and zero or more output artifacts.
The following is the high level structure of the build component:
{
"archives": [],
"drone_dimensions": [],
"gclient_variables": {},
"gn": [],
"name": "host_debug",
"generators": [],
"ninja": {},
"tests": []
}Each build element will be translated to an independent sub-build and its entire out directory will be uploaded to CAS.
gn, ninja, generators and tests properties are optional. Gn and
ninja properties can be used without generators or tests. Generators with
no gn and ninja properties is also supported.
An archive component is used to tell the recipes which artifacts are generated by the build and where to upload them.
By default the build output is archived to CAS in order to be used
as a dependency for global tests. If no CAS archive
is required cas_archive": false, needs to be added to the
configuration.
{
"name": "host_debug",
"base_path": "out/host_debug/zip_archives/",
"type": "gcs",
"include_paths": [
"out/host_debug/zip_archives/linux-x64/artifacts.zip"
],
"realm": "production"
}Description of the fields:
- name: - by default the entire build output is uploaded to CAS.
nameis used to associate the CAS hash to a value that can be referenced later as a dependency of global tests. Name is also used to select the folder from within src/out to upload to CAS. e.g if the build generates src/out/host_debug name must behost_debug. - base_path: - the portion of the path to remove from the full path before uploading to its final destination. In the example the above the base_path “out/host_debug/zip_archives” will be removed from the include path "out/host_debug/zip_archives/linux-x64/artifacts.zip" before uploading to GCS, e.g. <bucket>/flutter/<commit>/linux-x64/artifacts.zip.
- Type: - the type of storage to use. Currently only “gcs” and “cas” are supported. "gcs" uploads artifacts to GCS and "cas" to archive to CAS service. Cas value is used during development where we need to inspect the generated artifacts without worrying about location or cleanups. Gcs is expected for any artifacts being consumed by the flutter tool.
- Include_paths: - a list of strings with the paths to be uploaded to a given destination.
- cas_archive - a boolean value indicating whether the build output will be archived to CAS or not. The default value is true.
- realm - a string value of either
productionorexperimentalwhere production means the artifact will be uploaded to the location expected by the flutter tool and experimental will add anexperimentalprefix to the path to avoid interfering with production artifacts.
A list of strings with key value pairs separated by an equal sign. These dimensions are used to select the bot where the sub-build will be running.
To find the list of valid keys and values you need to select a bot from the
swarming UI.
On the dimensions section the left column contains the keys and
the right column contains the allowed values. If multiple values are allowed
for a key they are separated using | (pipe symbol).
"drone_dimensions": [
"device_type=none",
"os=Linux"
]In the previous example, the build containing this drone_dimensions component will run on a bot with a Linux OS that does not have any devices attached to it.
Drone dimensions accept values separates by | to specify more than one value
for the dimension. E.g. assuming the pool of bots have Ubuntu and Debian bots
a dimension of "os": "Debian|Ubuntu" will resolve to use bots running either
Debian or Ubuntu.
A dictionary with variables passed to gclient during a gclient sync operation. They are usually used to add or remove gclient dependencies.
"gclient_variables": {
"download_android_deps": false
}The example above is used to avoid downloading the android sdk dependencies in builders that do not need it.
A list of strings representing flags passed to the tools/gn script. The strings can be in the form of “--flag=value” or “--flag” followed by “value”.
"gn": [
"--runtime-mode",
"debug",
"--prebuilt-dart-sdk",
"--build-embedder-examples"
],The previous example will prepare the configurations to build a host debug version using a prebuilt dart sdk and also build the embedder examples.
A dictionary with two keys: “config” which references the configs created by gn and “target” which is a list of strings with the Ninja targets to build.
"ninja": {
"config": "host_debug",
"targets": [
"flutter/build/archives:artifacts",
"flutter/build/archives:embedder",
]
},In the example above the ninja command will use the configuration for host_debug and will build artifacts and embedder targets as described by the flutter/build/archives/BUILD.gn file.
This section of the build configuration is also known as local tests. It contains a list of dictionaries with configurations for scripts and parameters used to run tests inside the current build unit. These tests should not reference or use anything outside of the commit checkout or the outputs generated by running gn and ninja sections of the build configuration.
"tests": [
{
"language": "python3",
"name": "Host Tests for host_debug_impeller_vulkan",
"parameters": [
"--variant",
"host_debug_impeller_vulkan",
"--type",
"impeller-vulkan",
"--engine-capture-core-dump"
],
"script": "flutter/testing/run_tests.py",
"contexts": ["android_virtual_device"]
}
]Description of the fields:
- language - the executable used to run the script, e.g. python3, bash. In general any executable found in the path can be used as language. The default is empty which means no interpreter will be used to run the script and it is assumed the script is already an executable with the right permissions to run in the target platform.
- name - the name of the step running the script.
- parameters - flags or parameters passed to the script. Parameters
accept magic environment variables(placeholders replaced before executing
the test). Magic environment variables have the following limitations:
only
${FLUTTER_LOGS_DIR}is currently supported and it needs to be used alone within the parameter string(e.g.["${FLUTTER_LOGS_DIR}"]is OK but["path=${FLUTTER_LOGS_DIR}"]is not). - Script - the path to the script to execute relative to the checkout directory.
- contexts - a list of available contexts to add to the text execution step. The list of supported contexts can be found here. As of 06/20/23 two contexts are supported: "android_virtual_device" and "metric_center_token".
The test scripts will run in a deferred context (failing the step only after logs have been uploaded). Tester and builder recipes provide an environment variable called FLUTTER_LOGS_DIR pointing a temporary directory where the test runner can place any logs|artifacts needed to debug issues. At the end of the test execution the content of FLUTTER_LOGS_DIR will be uploaded to Google Cloud Storage before signaling the pass | fail test state.
Contexts are free form python contexts that communicate with the test script through environment variables. E.g. metric_center_token saves an access token to an environment variable "token_path" for the test to access it.
Note that to keep the recipes generic they don’t know anything about what the test script is doing and it is the responsibility of the test script to copy the relevant files to the FLUTTER_LOGS_DIR directory.
Generators are scripts used to generate artifacts combining the output of two or more sub-builds. The most common use case is to generate universal binaries for Mac/iOS artifacts.
Generators can be written in any language but they are required to follow some guidelines to make them compatible with the engine build system.
The guidelines are as follows:
- Flags receiving paths to resources from multiple sub-builds need to use paths
relative to the checkout (
src/) directory. If there are global generators in a build configuration, the engine_v2 recipes will download the full sub-build archives to the src/out/<sub-build name> directory. - Flags receiving paths to output directories must use paths relative to the src/out folder. This is to be able to reference the artifacts in the global archives section.
- The script is in charge of generating the final artifact, e.g. if the script generates multiple files that will be zipped later, then it is the responsibility of the script to generate the final zip.
- If the generator is producing a Mac/iOS artifact, then it is the responsibility of the script to embed the signing metadata.
Generators contain a single property “tasks” which is a list of tasks to be performed.
"generators": {
"tasks": []
}The example above represents a generator configuration with an empty list of tasks.
A task is a dictionary describing the scripts to be executed.
The property's description is as follows:
- Name - the name of the step running the script.
- Parameters - flags passed to the script. Both input and output paths must be relative to the checkout directory.
- Script, the script path relative to the checkout repository.
- Language, the script language executable to run the script. If empty it is assumed to be bash.
{
"name": "Debug-FlutterMacOS.framework",
"parameters": [
"--dst",
"out/debug",
"--arm64-out-dir",
"out/ios_debug",
"--simulator-x64-out-dir",
"out/ios_debug_sim",
"--simulator-arm64-out-dir",
"out/ios_debug_sim_arm64"
],
"script": "flutter/sky/tools/create_full_ios_framework.py",
"language": "python3"
}Tests in this section run on a separate bot as independent sub-builds. As opposed to tests running within builds, global tests have access to the the outputs of all the builds running in the same orchestrator build. A use case for global tests is to run flutter/framework tests using the artifacts generated by an specific engine build.
Global tests currently support two different scenarios:
- flutter/flutter tests with tester recipe. This workflow checks out flutter/flutter to run any of the existing sharded tests using the engine artifacts archived to GCS.
- complicated engine tests that require the outputs from multiple subbuilds with tester_engine. This workflow checks out [flutter/engine] and operates over the dependencies passed to it using cas.
Note: the supported scenarios can be later extended to support running devicelab tests although a smart scheduler is a prerequisite for it to be scalable(build/test separation model).
Framework test example:
{
"tests": [
{
"name": "web-tests-1",
"shard": "web_tests",
"subshard": "1",
"test_dependencies": [
{
"dependency": "chrome_and_driver",
"version": "version:111.0a"
}
]
}
]
}The property's description is as follows:
- name the name that will be assigned to the sub-build.
- shard the flutter/flutter shard test to run. The supported shard names can be found on the flutter framework test.dart.
- subshard one of the accepted subshard values for shard. Sub-shards are defined as part of the shard implementation, please look at the corresponding shard implementation to find the accepted values.
- test_dependencies a list of dependencies required for the test to run.
Engine test example:
{
"tests": [
{
"name": "test: lint android_debug_arm64",
"recipe": "engine_v2/tester_engine",
"drone_dimensions": [
"device_type=none",
"os=Linux"
],
"dependencies": [
"host_debug",
"android_debug_arm64"
],
"tasks": [
{
"name": "test: lint android_debug_arm64",
"parameters": [
"--variant",
"android_debug_arm64",
"--lint-all",
"--shard-id=0",
"--shard-variants=host_debug"
],
"max_attempts": 1,
"script": "flutter/ci/lint.sh"
}
]
}
]
}The property's description is as follows:
- name the name to assign to the sub-build.
- recipe the recipe name to use if different than tester.
- drone_dimensions a list of strings with key values to select the bot where the test will run.
- dependencies a list of build outputs required
by the test. These build outputs are referenced by the name of build
generating the output. This type of dependency is shared using CAS and
the contents are mounted in checkout/src/out. E.g. a build configuration
building the
host_engineconfiguration will upload the content of checkout/src/out/host_engine to CAS and a global test with ahost_enginedependency will mount the content of host engine in the same location of the bot running the test. - tasks a list of dictionaries representing scripts and parameters to run them.
Example task configuration:
{
"name": "test: lint android_debug_arm64",
"parameters": [
"--variant",
"android_debug_arm64",
"--lint-all",
"--shard-id=0",
"--shard-variants=host_debug"
],
"max_attempts": 1,
"script": "flutter/ci/lint.sh"
}The property's description is as follows:
- name the name assigned to the step running the script.
- parameters a list of parameters passed to the script execution.
- max_attempts an integer with the maximum number of runs in case of failure.
- script the path relative to checkout/src/ to run.
Global generators follow the same format as local generators but defined at the build top level. The main difference is that global generators can create new artifacts combining outputs of multiple sub-builds.
The archives component provides instructions to upload the artifacts generated
by the global generators. Is a list of dictionaries with three keys: source and
destination, and realm. source is a path relative to the checkout repository,
destination is a relative path to <bucket>/flutter/<commit>, and realm is
a string with either production or experimental value.
The realm value is used to build the destination path of the artifacts.
production will upload the artifacts to the location expected by the flutter
tool and experimental will add experimental as a prefix to the path to avoid
interfering with the production artifacts.
"archives": [
{
"source": "out/debug/artifacts.zip",
"destination": "ios/artifacts.zip",
"realm": "production"
},
{
"source": "out/debug/ios-objcdoc.zip",
"destination": "ios-objcdoc.zip",
"realm": "experimental"
}
]The example above will cause the file <checkout>/out/debug/artifacts.zip to be uploaded <bucket>/flutter/<commit>/ios/artifacts.zip.
Global generators can run locally if all their sub-build dependencies are downloaded. This section explains how to triage a local generator.
The instructions on this section can be used to triage problems with artifacts
created by glocal generators(E.g.Debug|Release|Profile-ios-Flutter.xcframework)
using the build outputs of CI subbuilds. During the migration to engine v2 we had
a regression in the size of the flutter libraries, using this process we were able
to inspect the files as they were generated by the CI, make changes to the generators
and run the generators locally to validate the fixes.
CAS client is required to download the sub-build artifacts. To install it in your machine run the following steps:
mkdir $HOME/tools- Download and unzip CAS binaries from [https://chrome-infra-packages.appspot.com/p/infra/tools/luci/cas]
- Add $HOME/tools to path and your ~/.bashrc
Create a gclient checkout following instructions from Setting up the engine environment.
CAS sub-build artifacts can be downloaded using information from a LUCI build.
Using [https://ci.chromium.org/p/flutter/builders/prod/Mac%20mac_ios_engine/2818]
as an example the execution details from steps 13 - 17 show the commands to
download the archives. -dir parameter needs to be updated to point to the
relative or full path to the out folder in your gclient checkout.
These are the commands to execute for our example build:
pushd <gclient checkout>/src/out
cas download -cas-instance projects/chromium-swarm/instances/default_instance -digest 39f15436deaed30f861bdd507ba6297f2f26a2ff13d45acfd8819dbcda346faa/88 -dir ./
cas download -cas-instance projects/chromium-swarm/instances/default_instance -digest bdec3208e70ba5e50ee7bbedaaff4588d3f58167ad3d8b1c46d29c6ac3a18c00/94 -dir ./
cas download -cas-instance projects/chromium-swarm/instances/default_instance -digest d19edb65072aa9d872872b55d3c270db40c6a626c8a851ffcb457a28974f3621/84 -dir ./
cas download -cas-instance projects/chromium-swarm/instances/default_instance -digest ac6f08662d18502cfcd844771bae736f4354cb3fe209552fcf2181771e139e0b/86 -dir ./
cas download -cas-instance projects/chromium-swarm/instances/default_instance -digest 1d4d1a3b93847451fe69c1939d7582c0d728b198a40abd06f43d845117ef3214/86 -dir ./The previous commands will create the ios_debug, ios_debug_sim, ios_debug_sim_arm64, ios_profile, and ios_release folders with the artifacts generated by its corresponding sub-builds.
Once the checkout and dependencies are available locally we can cd|pushd to the root of the client checkout and run the global generator. The command can be copied verbatim from the executions details of the build.
The following example will run the generator to create the ios artifacts:
python3 flutter/sky/tools/create_full_ios_framework.py --dst out/release \
--arm64-out-dir out/ios_release --simulator-x64-out-dir out/ios_debug_sim \
--simulator-arm64-out-dir out/ios_debug_sim_arm64 --dsym --strip