|RFC-0093: Design principles of component manifests|
Design notes on the `.cml` and `.cm` formats used for component manifests.
|Date submitted (year-month-day)||2021-04-26|
|Date reviewed (year-month-day)||2021-05-05|
This document captures design deliberations, principles, and decisions made
.cm formats used for
Most of the decisions described below were made in the years 2018-2019.
- Component manifests have a frontend and backend
- A component is defined by a manifest
- Component manifests are declarative
- Component manifests can come from various sources
.cmlis the common filename extension for component manifest source files. CML files are a JSON5 DSL for declaring a component.
- ComponentDecl is a FIDL table which is the canonical wire and storage format for component manifests.
.cmis the common filename extension for files that contain a ComponentDecl in the FIDL envelope binary format.
cmcis a command line host tool for generating
.cmlfiles. It is built in the Fuchsia source tree and distributed as a prebuilt executable through the Fuchsia SDK.
#1: Component manifests have a frontend and backend
Humans and machines have different needs and preferences. In designing the component manifest syntax and formats, a key design principle is to create a single backend format for the Component Framework to read, and a separate frontend (singular at first) for developers to write. This design decision has the following benefits:
- It is possible to optimize the backend and frontend(s) separately to satisfy different design goals.
- It is possible to evolve the frontend without modifying the backend, and vice versa.
- Although currently there exists only one frontend, it is possible to introduce additional frontends, such as to satisfy different design goals for different audiences or to cater to preferences of convenience, familiarity, or style.
To support these goals, the SDK provides
[cmc] ("component manifest
compiler"), the standard tool to convert component manifest source files
.cml) to manifest binary files (
cmc is usually integrated
transparently with the build system, which means that developers normally
interact with source files unless they are debugging or performing analysis on
ComponentDecl: the component manifest backend
ComponentDecl is the canonical storage and wire format for component
manifests. It is designed to be written and interpreted by code such as
component manager, component resolvers and manifest
analysis tools such as
fx scrutiny. Such a format needs to meet
the following goals:
- It must be unambiguous and self-describing. It should be possible to directly derive the meaning of a manifest from its contents.
- It must be able to evolve over time, supporting forward and backward compatibility.
- It must be straightforward to parse and avoid gratuitous format conversions. Otherwise it would needlessly increase the risk for bugs or attacks in code that handles component manifests, including component manager.
- It must be easy to integrate with the runtime, components and tools that interface with component manifests.
These design goals led to a natural choice of format: FIDL. FIDL is the standard wire format for IPC in Fuchsia. FIDL value types (i.e. types with no handles) can be persisted and used as a storage format. Specifically FIDL envelopes are used because they have the added benefit that they skip over unknown fields, which is useful for backward and forward compatibility. Since FIDL bindings are already present for any runtime that runs on Fuchsia, it is easy to integrate with code and requires no additional support for parsing or conversion.
ComponentDecl is structured in such a way that there are no defaults: in other words, a field is only unpopulated if it is not applicable or it came from a version of the manifest that did not have that field. Also, to support forwards and backwards compatibility, ComponentDecl and the structures nested in it are FIDL tables or FIDL flexible unions.
CML: the component manifest frontend
CML ("component manifest language") is the source format for component manifests. It is designed to be read and written by humans, but may also be read and written by development tools such as formatters and language servers.
- It should be readable by a novice who understands basic Component Framework concepts.
- It should be convenient to represent common patterns without excessive boilerplate.
- Changes that affect the syntax but not the semantic meaning of the manifest should be possible without requiring any changes to the manifest's binary representation. For example, a change from a singleton array to a singular value should not affect the output.
- It should promote maintainability. In particular, it should allow comments.
- It should be machine-friendly enough to support automated transformations, for example to support large scale refactors.
CML was invented to meet these goals. CML is a JSON5-based configuration language which acts as a simple DSL which generates ComponentDecl. By using JSON5, CML leverages a language that is already familiar to many developers and is widely used elsewhere in Fuchsia. Unlike ComponentDecl, CML provides some affordances that make it possible to write manifests more succinctly:
- It allows defaults to be elided for some fields.
- It allows multiple capabilities to be grouped into a single declaration as long as they share the same options.
- It allows manifests to include manifest shards that contribute contents to the manifest. For example, you can depend on a library and include the library's shard to get all the capabilities required by that library.
Finally, the translation from CML to ComponentDecl, while not one-to-one, should be straightforward for users to understand without having to learn the rules.
#2: A component is defined by a manifest
A component is described by a manifest. The manifest is resolved at start time
via the URL of the component. For example, a component launched via a
fuchsia-pkg:// URL with will have a manifest with the
extension, containing a serialized
ComponentDecl, that is resolved
from a package. In addition to the manifest, a component may also incorporate
resources from the same package. For example, a component that uses the ELF
runner specifies the location of the ELF binary in that
package. On the other hand, a component with an
https:// URL may have a
ComponentDecl that is generated by the
https resolver but is not backed by a
resource obtainable via a URL.
A component's manifest fully describes its inputs, outputs, and internal composition. Currently, component manifests cannot have any parameters or "dangling" values that are filled in at runtime.[^1]
That isn't to say a component's behavior is completely described by the manifest, however. For one, the capabilities offered to a component are determined by the parent; the component has no control over who provides those capabilities. Also, every component is part of an environment which provides certain types of configuration for the component, such as what resolvers are used to resolve component URLs.
Following the URLs between manifests yields a
component instance tree. The component instance tree is a
comprehensive description of the software that constitutes a Fuchsia image. This
makes it possible to perform security auditing with confidence over a given
system image, such as with
#3: Component manifests are declarative
While configuration languages with imperative features are powerful, they come at cost of lost readability, predictability, and auditability. Precedent shows that configuration languages with too much imperative flavor are brittle and less user-friendly.[^2] In Component Framework's case, component definitions must be auditable and readily understandable, which makes an imperative-style configuration language a non-option.
Thus, CML is a declarative language. To the extent that CML supports generating parts of the manifest, this is only supported in cases that have very predictable outcomes. For example, manifests support default values and inclusion, but they do not provide templatizing or parameterization features.
CML is a language designed to be read and written by humans. With the exception of developer tooling integration (for example, formatting tools or IDE templates), CML is not intended to be generated by tools. Generating CML files carries an elevated risk of obscuring the underlying contents of the manifest, since there are now three layers involved: CM, CML, and the tool. If for some reason a manifest has to be generated, you should write a separate frontend to generate CM.
#4: Component manifests can come from various sources
Component manifests, in general, are not bound to any single distribution
mechanism. It is ultimately the responsibility of the
component resolver to retrieve the component manifest for
a URL. How a resolver accomplishes this is particular to a given URL scheme. For
fuchsia-pkg:// resolver will retrieve the package and read the
manifest from it designated by the fragment identifier part of the URL. A web
resolver might generate a manifest whose contents may vary based on domain,
security policy, and user preferences.
Currently, the most common way to distribute a component is through a Fuchsia
package. Such components are identified by a
fuchsia-pkg:// URL. The
component's manifest is shipped as a blob in this package, usually in
- Declarative application management in Kubernetes: principles used in designing the configuration language for Kubernetes, and a study of alternatives.
- Imperative vs declarative: expands on the titular topic.
- Starlark: a Python-based DSL and an imperative configuration language.
- Jsonnet: an extension of JSON into a data templating language, where any program produces a JSON document.
- borgcfg, GCL, and borgmon: Functional configuration languages used by Google, roughly analogous to Kubernetes, whose history helped inform us on the tradeoffs between imperative and declarative syntax.
[^1]: In the future, there is a high probability that manifests need to support some sort of parameterization feature to support variants and product configurability. When we do so, we should approach this in a way that avoids the common pitfalls associated with parameterizable configurations.
[^2]: For more context on this point, Kubernetes has extensive documentation explaining many of the downsides of non-declarative configuration.