|RFC-0095: Build and assemble Workstation out-of-tree|
Creates a workflow which allows users to build and assemble the Workstation product outside of fuchsia.git.
|Date submitted (year-month-day)||2021-03-24|
|Date reviewed (year-month-day)||2021-05-19|
Note: This RFC has been superseded by RFC-0220, which deprecated the workstation product. This RFC is preserved for posterity, but we will no longer pursue implementation.
The Workstation product, defined as the workstation_session, ermine, terminal and simple browser, is currently built and assembled in the Fuchsia source tree. There is no clear distinction between Fuchsia as a platform and the workstation product. This requires that the platform and the product must be built at the same time to assemble the final product. We propose that this coupling be broken to allow the Workstation components to be built separately from the platform and that the final product be assembled outside of the Fuchsia source tree. This proposal does not address removing the implementation of the Workstation product from the Fuchsia source tree but rather focuses on the dependent problem of out-of-tree product configuration and assembly. This work will lay the groundwork needed to eventually move the Workstation product out-of-tree completely but that is out of the scope of this document.
It is important to call out that the terminal is considered to be a fundamental component for the Workstation product, however, the terminal is used by other products outside of the Workstation product. Therefore the development of the terminal component will remain in fuchsia.git and it will be consumed by the Workstation as part of the platform until a further decision is made regarding the future of this component.
There is currently no way to build and assemble a product on Fuchsia without building the entire Fuchsia platform at the same time. We can build individual components outside of fuchsia.git using the SDK but those artifacts need to flow back to Fuchsia Global Integration to be assembled into a final product. We would like to be able to offer stable versions of the platform for products to be built on without needing to have access to Fuchsia Global Integration.
There are still many unknowns when it comes to building products out-of-tree which is why we want to start with the Workstation product. The Workstation product is not a production-grade, supported user-facing product, it is a reference for developers and an environment for early adoption of platform features and testing, so there is a higher tolerance for slowdowns at this stage. The Workstation product is also meant to be the environment in which enthusiasts can explore Fuchsia; by removing the need to build the entire platform we can make it easier for developers to play with Fuchsia.
Moving Workstation out-of-tree builds on top of the Standalone Image Assembly Tool RFC-0072, this proposal assumes that this tool exists.
The source code for the Workstation repository will be hosted on git-on-borg. Hosting the code on git-on-borg will allow us to use existing Fuchsia infrastructure and tooling which will allow us to focus on the task of building and assembling the product.
For the remainder of this document workstation.git will refer to the repository hosting the code for the Workstation product.
The focus of this project is to build and assemble the Workstation product out-of-tree and thus we will not be moving the development of the components out of fuchsia.git until we have worked out how to do this. Since the component development will remain in-tree, we can rely on the existing test and build infrastructure to verify that the Workstation product does not regress. The process of rolling new versions of the platform, sdk, experiences.git and other dependencies will be done manually with a script to allow the team to focus on the problem at hand. When we begin to transition the development of the component out-of-tree we will need to revisit how we do continuous integration.
The daily builds of the platform will need to be automated but we can use the existing Fuchsia infrastructure. A daily builder will need to be created which builds the packages required to assemble the Workstation product and upload those artifacts to some storage repository. The technical design for this workflow still needs to be figured out. We will consult with infrastructure and security before we start working on this.
Building the Workstation product, as opposed to core, will ensure that we have all the artifacts that we need for product assembly even if we do not build them at this point. The artifacts that we do build out-of-tree will be used in place of the artifacts that rolled with the platform. Doing this will also allow us to use the existing Workstation builder which will minimize the amount of extra infrastructure we will need to manage.
The Workstation repository will use a combination of git submodules and Bazel workspaces for dependency management. Git submodules will be used to pull in external repositories whereas Bazel will be used to download prebuilts and toolchains required by the build.
Bazel's toolchains will be used in conjunction with workspace rules to download the appropriate prebuilts that are required for their respective build rules. One limitation at this point is that many of our prebuilts are stored in CIPD and require the cipd command line tool. Initial prototypes require that the cipd tool be somewhere on the developer's path. We are actively looking into alternaives and do not think this will be a long term requirement.
The directory structure for the Workstation repository will closely resemble the fuchsia.git directory structure to keep consistency for Fuchsia developers. The root directory will contain various metadata files needed for the project as well as the following top-level directories.
- //src - Source code used to build the Workstation product.
- //tools - scripts and tooling to support the build.
- //src/experiences - the existing experiences.git repository.
- //prebuilt - any prebuilts that are needed.
- //third_party - third party code that is used by the src directories.
The Workstation product will be built with the Bazel build system. The decision to use Bazel diverges from fuchsia.git which uses gn/ninja. This decision means that we will not be able to lean on the expertise that the Fuchsia team has acquired or reuse the existing build rules. However, the current gn SDK does not contain any logic for building or testing Flutter applications which is the majority of the Workstation product. The team would need to write these rules no matter what build system they choose. The sample Bazel SDK that exists in-tree does have some rudimentary logic for building Dart and Flutter applications that we can build off of.
The Workstation team has chosen to use Bazel over gn/ninja for a few main reasons that they feel makes this a more attractive build system to use.
open source adoption: Bazel is not the most popular build system but it does have more adoption than gn/ninja and thus should feel more familiar to the open source community.
ecosystem of rules: Bazel has a rich ecosystem of existing rules which help to promote reuse.
dependency management: Bazel can manage external dependencies.
Fuchsia SDK Bazel SDK productionization: Using Bazel for the Workstation product gives us the opportunity to push changes into a Fuchsia Bazel SDK which can be used by other projects.
single build/analysis phase: A common complaint for developers coming to Fuchsia is that they do not know why a package or tool is not included in their build. This often comes down to a user not including the target in their gn args. Bazel removes this failure mode by requiring the user to explicitly specify what they want to build, run or test.
hermetic builds: Bazel has hermetic builds by default whereas gn does not.
During this phase we will need to write build rules to build the experiences repository against the workstation.git buildroot. We will utilize the new_local_repository feature of Bazel. Using this rule will allow us to keep the Bazel specific build rules in workstation.git to build against sources mapped in from the experiences repository. It is important to note that this setup is prone to build breakages because the Bazel build rules are hosted in a separate repository from the actual source code. These breakages will only affect the team that is working on getting out-of-tree assembly and not any in-tree developers. We are aware of the trade-off here but feel it is better to disrupt the workflow of the out-of-tree developers instead of the in-tree developers. The out-of-tree developers will be responsible for fixing the breakages as they occur.
new_local_repository( name = "ermine", path = "vendor/experiences/session_shells/ermine/shell/", build_file = "src/ermine/BUILD.ermine", )
flutter_component(name = "ermine_component", ...) flutter_package(name = "ermine", deps = ":ermine_component")
The Workstation components are currently written in Dart and Rust of which Dart is the only language that is currently supported for out-of-tree development. The workstation_session and terminal components are written in Rust and Ermine and simple-browser are written in dart. We will start by focusing our attention on building the Dart components while continuing to build the Rust components in-tree and delivering them as part of the platform.
There are currently two blockers for building the workstation_session out-of-tree. The first is that the workstation_session is using several private APIs which means that it could not be built out-of-tree even if we had the language support. The second blocker is the lack of out-of-tree Rust support. The workstation_session is currently being broken up into smaller platform level components which is eliminating its dependency on private interfaces and out-of-tree Rust support is being explored. We will make a decision about the fate of the workstation_session as these two projects progress. If the workstation_session is completely migrated off of private interfaces but Rust support is not going to come soon after then we will rewrite the session in C++ whereas if Rust support lands before or soon after then we will keep the session written in Rust.
The fuchsia.git tree supports various sanitizers for both local development and in the infrastructure builders. Support for these system in the workstation.git repository is not an immediate goal since we will not be moving source code out of fuchsia.git until we have proven that the build and assembly system works with the mapped in repository. However, we should block on moving code out of fuchsia.git until we have sanitizer support since it introduces a regression from the current setup.
When we get local builds working properly we will add support for passing build time flags to enable the various sanitizers locally. This will allow us to build against the mapped in sources with sanitizers enabled. When we start moving code out of fuchsia.git we will set up builders to match the current experience of committing to fuchsia.git.
Fuchsia code can currently be searched on code search. The workstation.git repository will be added to this same project as a separate project which will allow it to be searchable and viewable online.
We do not plan to enable code coverage for workstation.git at the beginning of the project and do not have any plans for the near term. The main reason for this is that the majority of the Workstation code is written in Dart which does not currently have sufficient enough code coverage support to warrant us investing the time needed to enable it.
We will be losing parity with the current experience for writing c++ code since it does show incremental code coverage but the scope of work to enable this in workstation.git is too large to take on at this point. However, it does present a good opportunity to work with the Fuchsia build team to make these tools more readily available for out-of-tree developers.
Third Party Support
All of the source code written in workstation.git will need to share the same third party libraries. The tooling to support this has been solved many times over in other large repositories so we will not have to start from scratch but it does represent a quantity of work that we will have to do and support for the entirety of the project.
Licensing compliance will match what is used by the Fuchsia Project. Compliance for third_party repositories will be monitored by the top-level OWNERS of the repository as part of the code review process which brings these dependencies in.
We currently do not have any solid plans for checking licenses and generating NOTICE files during builds and during code checks. We will need to study how this is done in the fuchsia.git repository and port that logic to this repository.
Binary Size Monitoring
Monitoring the size of our binaries is important for the long term health of our project. There exist tooling in the Fuchsia build system which monitors the size of binaries that are produced during each commit. Building these systems again for the Workstation product is out of scope for the early phases of this project but it is something we will need to add eventually.
We need to investigate how tightly coupled these systems are with the current fuchsia.git build system to see if we can simply integrate with them in their current forms or if we will need to do work to decouple them. If we need to work to decouple the tooling from the build system we should focus on making these tools reusable to scale out to other product integrators in the future.
There is currently a project taking place to add performance monitoring to the Workstation product. This project is already proceeding with a focus on making the monitoring available to out-of-tree users by only using tooling that is available in the public SDK. We plan to move these tests into the workstation.git repository as the code moves out of fuchsia.git.
The initial set of contributors to the workstation.git repository will be small compared to the set of contributors to fuchsia.git. For this reason, enforcement of OWNERS files will not be a priority during the initial phases of this project but we will continue to add them to the appropriate directories to facilitate communication. As we start to move code out of fuchsia.git we will introduce OWNERS files and enforce the same rules for code reviews that we do in fuchsia.git.
Contributions to the workstation.git repository will be governed by the same policies that govern contributions to fuchsia.git since they will be hosted under the same project.
The repository will use the existing Fuchsia Monorail bug tracker to track bugs. The reason for using Monorail is because it is already set up for tracking Fuchsia bugs and because the board support for the Workstation product will still remain in fuchsia.git. There is a top-level Workstation component which can be used for general triage and then there are several existing components for tracking the various features of the Workstation product.
Flow of Artifacts
In order to assemble the Workstation product outside of fuchsia.git we will need to provide prebuilts of the Fuchsia platform to be consumed by the Workstation product. This implies that we will have two classes of artifacts, the artifacts that comprise the Platform and then the artifacts that are layered on top to assemble the final Workstation product image. The Platform images will be built ahead of time in to be consumed by the workstation.git repository whereas the Workstation components will be built at the time that the final product image will be assembled. The set of artifacts that make up the Platform is not clearly defined at this point as the Platform is constantly evolving but the Workstation components will include the workstation_session, Ermine and simple-browser. It should be noted that we do not propose to move any driver development out-of-tree as part of this proposal even if it is a driver that only supports the Workstation product.
To allow workstation.git to consume the prebuilt artifacts we will need to set up a builder of fuchsia.git to create the prebuilts and an autoroller which will pull the artifacts into workstation.git. We may be able to extend one of the builders which is currently running today but we have not fully investigated if that is possible.
The fuchsia.git builder will run once a day and will build the Fuchsia platform. Exactly how this builder will be configured is yet to be determined because we need to understand exactly what gets built. The builder will upload all of the appropriate artifacts to a location that the autoroller can consume. The exact location and mechanism for uploading these artifacts is still to be determined. Included with the artifacts will be a manifest file which includes all of the dependency versions used to create the platform. This includes, but is not limited to, the SDK version, the version of experiences.git, the Flutter and Dart versions, the Fuchsia toolchain version, etc.
The autoroller will be run multiple times a day to check for new versions of the platform. Running multiple times will ensure that we pick up the most recent versions regardless of flaky rolls or builds that may come at different times throughout the day. The autoroller will poll the latest version of the platform and pull in those artifacts. The roller will also read the manifest file which contains all of the dependent versions and update those as well. This will ensure that we do not have version skew between the platform and the various dependencies.
The required data are mostly files containing file system partitions, such as fuchsia.zbi and fvm.sparse.blk. These files contain all the data necessary to boot, as well as all packages in the ‘base’ and ‘cache’ lists. These files will be collected into a .far “product package”, and the metadata found at the URL of this package will contain all necessary metadata to build component packages with a matching SDK. Packages built against this metadata can be appended to base and cache by modifying the system partitions contained in the “product package”. There is parallel work for creating a standalone out-of-tree image assembler, so the list of specific artifacts that are bundled in the “product package” is subject to change depending on the API for the new tool.
The daily builder will upload the artifacts in a way that will allow us to access the "product package" by a given version. A fixed repository will contain references to all versions of the platform, to enable the roller to identify the latest build or a specific build suitable for ABI compatibility. The specific process which will enable this is currently being explored external to this proposal.
When assembling the final product, the build system should first download the “product package”, which is identified by URL. It will then download a matching SDK. (The SDK could also be bundled within the “product package”). It should build packages against that SDK, creating a meta.far file for each package, as well as collecting all necessary blobs into a single location. The assembly tool will then append these blobs and packages to the “product package”, resulting in a completed product build. The resulting artifacts should also have all necessary information to generate OTA packages.
Compiling the Workstation product separately from the platform exposes us to potential ABI incompatibility issues. We need to make sure that we are always compiling the Workstation components with an SDK release version that is either equal to or older than, but within the compatibility window per RFC-0002, the release version of the platform that the Workstation product is being assembled with. In order to ensure that we maintain this invariant, we will include the SDK version in metadata that rolls along with the platform and the workstation.git repository will be updated to use this version of the SDK.
It should be noted that this ABI consideration needs to be extended to other petals that contribute code to the Workstation product, like the Flutter and Dart runners. To minimize the risk of ABI breakages coming from these petals we will initially roll them as part of the platform which means that they will have already been validated in fuchsia’s global integration.
As we start to move petals out of the platform surface we will need to make sure that they are compatible with our product. For petals that come back into Fuchsia Global Integration we can propagate metadata about the versions as we build the platform itself. However, there is a future in which some of our dependencies do not come into Fuchsia Global Integration which means we will not have version information included in the platform rolls. We will need to use platform versioning to make sure we are including the correct versions and we will need to work out a way to handle out dated versions.
The process to build and assemble the Workstation product out of fuchsia.git will take place over the course of several phases, each one building on the previous. This phased approach is required since pieces of the system which support out-of-tree development are currently in progress.
The initial phase for getting the Workstation to build out-of-tree is to set up the development environment. We will create the workstation.git repository to host the code. This repository will contain a bootstrap script which will allow users to get up and running. The repository will use Bazel workspaces and git submodules to manage third_party code, prebuilts and the vendored in experiences repository.
The second phase will focus on getting the build system set up to build the Flutter components. We will use the existing Bazel rules that live in fuchsia.git as a starting point to write the build rules. The end state will be when we can successfully build a Flutter component that can be launched on an already running device.
At the same time as setting up the rules for building the Flutter components we will explore how to do assembly out-of-tree using the product assembly tool. This process will be done manually by just copying all of the artifacts that are built in fuchsia.git into the workstation.git repository’s out directory and trying to create a bootable image. The end state will be when we can provision a device with an image that is assembled out-of-tree but built in tree.
The third phase of this transition is when we work towards bringing all of these pieces together to assemble the entire product. This includes setting up all of the infrastructure required to build the platform artifacts and upload them to the TUF repository on a scheduled basis. This builder will run in Fuchsia infrastructure. The workstation.git repository will roll the new version with an autoroller. The autoroller will update a manifest file in the workstation.git repository, run integration tests and commit this change. This differs from fuchsia.git which commits version updates to an integration repository to coordinate version changes.
The workflow for building and assembling will be:
- Roll the new version of the platform by updating a manifest entry which identifies the platform version and all dependent versions. How the versions are read and identified needs to be designed.
- Use Bazel and git to download the appropriate prebuilts and third party repositories needed to build the in-tree components.
- Use the platform download tool to download all of the platform artifacts from the TUF repository.
- Build the in-tree components.
- Assemble the built components and the prebuilt components into a final image.
This project should not have any impact on performance of the Workstation product since it is only building and assembling. However, we should start to see significant improvements in the build times for Workstation developers since they will not need to build the platform.
We are reusing much of the same infrastructure that fuchsia.git uses but we need to ensure that Workstation is being built and released in a secure manner. We will work with the security and infrastructure teams to create a secure release process and build pipeline.
The Workstation product is only intended to be a reference product but we will want to make sure we are still following proper security protocols. Once we get closer to doing any sort of release we will need to work with security to ensure that we are building our releases in a security compliant way.
There is no impact on privacy to consider at this time. We will be using open tooling that does not collect user data.
The Workstation product has a suite of tests that are run on each commit. These tests include unit tests, integration tests, e2e tests and performance tests. The code for these tests will be ported to the workstation.git repository as the code for the Workstation components are ported to worksation.git.
The logic for these tests live in experiences.git so we will be able to run them since we are mapping this repository into workstation.git until we actually move the code. All of these tests are using code and tools that are available in the public SDK so we will be able to run them out-of-tree but will need to update our build system to ensure that they can be built and run outside of fuchsia.git.
The tests will be run on every commit to the workstation.git repository to prevent regressions. This will require collaboration with the infrastructure team to figure out how to build and run these tests in CQ. The current infrastructure assumes that fuchsia.git is present so we will need to develop a method that is build system and repository layout agnostic.
We will only need to add documentation to workstation.git which explains how to get the repository set up for development. The Workstation development will remain in-tree so we should continue to keep pointing our documentation to that workflow.
Drawbacks, alternatives, and unknowns
One alternative is to just do all of this work inside of the Fuchsia tree but using tooling that exists entirely in the sdk. We moved away from this approach because it makes it too easy to hide our usages of private interfaces. Moving the assembly completely out-of-tree forces us to rely just on the sdk.
There are still some unknowns around exactly what we need to upload to the TUF repository, the shape of the upload and the format of the metadata that we upload with the artifacts. We feel that these specifics will be identified as we start working through the process and we will amend this RFC as needed.