A typical model would be a tree that looks like this:
The build system output is used to construct one or more AndroidModel instances, which are the root of the model. Each AndroidModel instance contains one or more AndroidProjects. In the case of Gradle there would always be one AndroidProject per AndroidModel and in the case of Blaze there would be many.
Projects contain any number of Variants. Variants contain any number of Artifacts, and the artifact contain the majority of metadata for the project (application ID, etc.). Artifacts also crossreference a list of Configs, which serve a similar purpose to ProjectFlavorContainer -- they contain metadata and a list of pathnames that make up the source folders.
The IDE attaches exactly one AndroidModel instance to each module.
If the IDE wants to use the notion of a selected variant, it is the responsibility of the IDE to implement this concept. It is not represented in the model, which treats all Variants as equivalent.
Paths are represented using PathString objects. PathStrings are lightweight identifiers that can identify any io.File, nio.Path, or VirtualFile object that would be used by Android Studio. Unlike Path and VirtualFile, it doesn't depend on an actual filesystem implementation.
Typically, the paths provided by a real build system will all be files on the local filesystem that can be converted back to io.File. However, by using a more general path identifier in the model, unit tests can construct models in their choice of virtual filesystem. This generalization would also potentially allow us to refer to files within jars or in other custom filesystems.
In most cases, the paths provided by the model are just the places where the build system will look for an input file or write an output file. The fact that a path appears in the model does not guarantee that there is anything on the filesystem in that location, and clients must check for the file's existence if they want such guarantees.
Gradle creates one Module for each leaf Gradle project. Each Module has an AndroidModel containing a single AndroidProject (that is, the projects list in AndroidModel always has exactly one element).
There is a 1:1 mapping from Gradle AndroidProject instances and projectmodel AndroidProject instances.
Projectmodel doesn't expose the Gradle model version number. All features of different versions of the Gradle model are normalized during the conversion to projectmodel, and any feature that is missing from a specific Gradle model version will be exposed in projectmodel as a missing optional feature.
The project name and project type attributes map fairly directly from Gradle and projectmodel.
ConfigTable describes all the build types and flavors present in the project. It also describes the default configuration and all the source providers, including the main source provider and any per-artifact source providers.
Gradle projects always fill in the optional configTable attribute on AndroidProject. The first dimensions of the config table correspond to flavor dimensions, in the same order dimensions are declared in the build.gradle file. The last dimension corresponds to the build type.
Cells in the ConfigTable correspond to variants of a specific artifact. They are identified using an ArtifactVariantPath. ArtifactVariantPath are a lot like Gradle variant names, except:
For example, the variant path pro/release/main would identify the proRelease variant's “main” artifact (where proRelease is the variant that uses the pro flavor and release build type).
Each configuration (build type, flavor) and source provider pair is combined into an instance of Config. The Config describes all the source folders involved in that build type/flavor along with any configuration metadata it overrides.
Configs are inserted into the config table with a ConfigPath. The path identifies which variants and/or artifacts the Config should be used with. So from the perspective of a model consumer, a build type or flavor would both be instances of Config. The only difference is what artifacts the Config is attached to, and this is controlled by the path.
Build types are represented in the ConfigTable as a Config with a ConfigPath that only matches the second-last segment. So, for example, the “debug” build type in a 2-dimensional config table would use the path null/debug/null. The first null means that it applies to all flavors. The “debug” segment indicates that it applies to the debug build type, and the final null indicates that it applies to all Artifacts from that build type.
Flavors are also represented using Configs, except that they use a path that matches an earlier segment.
Flavors also differ from build types in that their Config will never override certain attributes - such as isDebuggable. Configs return null from such attributes, indicating that they don't modify that attribute.
Gradle permits custom configuration to be attached to specific flavor combinations. Such flavor combinations are also represented as Configs but they use a path that matches multiple segments. For example, the following path would match only match the v2 and pro flavors:
The default configuration for a project (the “main” source tree) is attached to the ConfigTable with a path that matches everything. It is always the first Config in the table, indicating that everything else overrides it.
Configuration and source files that only apply to a single artifact also use Configs, but they use a path that only matches the last segment. For example, this path would match all variants of the “main” artifact:
Gradle inserts Configs into the ConfigTable in priority order. That is, configs inserted later will override identical attributes in earlier configs.
Gradle artifacts map fairly directly onto the projectmodel Artifact class. However, the way API consumers access the artifact's metadata differs. In gradle, the AndroidArtifact class has a bunch of methods that return metadata (such as getApplicationId()).
Artifacts in projectmodel store the result of merging their configs (the main config, flavors, build type, and any per-artifact configuration) in a Config called “resolved”. Application code that wants to know what went into the artifact but doesn't care what flavor it came from will normally use Artifact.resolved rather than examining the config table.
So, for example, to access the application ID using a projectmodel Artifact, a caller would invoke artifact.source.overrides.applicationId.
In addition to the resolved config, each artifact has cross-references to the constituent configs that went into the artifact, in priority order. So - for example - a typical artifact might reference the config for the main source provider, one for the build type, one for the flavor, and one for a variant-specific override.
Gradle always populates the optional variantPath attribute of Variant. The last segment of the variantPath is the build type, and the preceeding segments correspond to flavors.
There is no merged flavor associated with Variants in projectmodel. API clients should use the source from one of the artifacts instead. (Doing so is more accurate than using the Variant info since it includes any per-artifact configuration and overrides).
Gradle will not fill in the optional dependency information for individual configs in the config table (since that information is not reported by the Gradle model), but it will fill in the resolved dependencies for each Artifact.
This section describes the mappings from blaze BUILD files to the projectmodel objects.
Blaze only creates one Module. It will have an AndroidModel containing many AndroidProject instances.
Each android_binary rule will become an AndroidProject of the APP type. Each android_library rule will become an AndroidProject of the LIBRARY type. Each android_manifest_merge rule will become a Config containing only a list of manifest files. Each android_resources file will become a Config containing only a list of resources. android_test and android_robolectric_test targets will also become AndroidProjects of type TEST.
The AndroidProject rules created by Blaze will have one Variant containing one Artifact. There may be multiple Configs attached to each Artifact, one containing all the source included in the blaze rule, along with separate Configs for the android_resources and android_manifest_merge rules that are referenced by the original rule. The name of the Configs will match the name of the blaze rule they came from.
Blaze will not supply a config table for its models.
Blaze will not fill in the dependencies for Configs that come from a android_manifest_merge rule or android_resources rule, but will fill them in for resolved configs and for the configs that come directly from the rules that produced AndroidProject instances.
Note that - unlike gradle models where each Config contains no more than one manifest - it will be common for Blaze models to contain multiple manifests per fragment.
Legacy intellij projects that don't use gradle or blaze will have a project with one Variant and two Artifacts (for test and app sources). The attributes will be extracted from the manifest and the IntelliJ module metadata.
This section describes the general design principles that guided the design of the model as a whole.
Features are added to the model as they are justified by use-cases in the build-system-independent tooling. The model is not required to be sufficiently complete to build the project, so some information known to the build system will not be exposed in the model.
The presence of features that aren‘t supported by all build systems or that aren’t present in all projects will be exposed explicitly via capabilities. API consumers are required to check before accessing that feature. For example, the presence of a capability might be indicated by a nullable field.
The entire model is invariant. The build system is required to replace the model's root when it changes, using shallow copies at its option.
Kotlin data objects are preferred for all data types. If portions of the model are found to be inefficient to construct for large models, those portions of the model can be lazily constructed provided the model types remain externally invariant and implement correct value-based hashCode and equals implementations.