Android Architecture

Android Architecture

Overview

Good architecture makes Android apps testable, maintainable, and scalable. This skill guides architectural decisions: module structure, layer separation (UI → Domain → Data), dependency injection with Hilt, state management with Coroutines and Flow, and patterns that survive configuration changes and process death.

When to Use

• Starting a new Android project or module
• Adding a feature module to an existing multi-module project
• Choosing between MVVM and MVI for a screen
• Setting up or refactoring dependency injection
• Defining boundaries between layers
• Debugging lifecycle or state management issues

Skip when: Making a small change within an already-architected feature.

Core Process

Step 1: Module Structure

1. Organize by layer and feature:

:app                          # Application, navigation, DI entry point
:feature:home                 # Home screen (UI + ViewModel)
:feature:profile              # Profile screen (UI + ViewModel)
:feature:settings             # Settings screen (UI + ViewModel)
:core:data                    # Repositories, data sources, API services
:core:domain                  # Use cases, domain models (pure Kotlin)
:core:database                # Room database, entities, DAOs
:core:network                 # Retrofit services, API models, interceptors
:core:ui                      # Shared composables, theme, design tokens
:core:common                  # Utilities, extensions, constants
:core:testing                 # Shared test utilities, fakes

2. Dependency rules:
   • :feature:* depends on :core:domain, :core:ui, :core:common
   • :core:data depends on :core:domain, :core:database, :core:network
   • :core:domain depends on nothing (pure Kotlin module)
   • :app depends on all :feature:* and :core:data (for DI wiring)
   • Never: feature → feature, core → feature, domain → data

Step 2: Layer Separation

3. Three layers, strict boundaries:

Layer	Contains	Depends On
UI	Composables, ViewModels, UiState	Domain
Domain	Use cases, domain models, repository interfaces	Nothing
Data	Repository implementations, data sources, mappers	Domain (interfaces)

4. Data flows one direction: UI ← Domain ← Data

// Domain layer: pure Kotlin, no Android dependencies
interface TaskRepository {
    fun getTasks(): Flow<List<Task>>
    suspend fun addTask(task: Task)
}

data class Task(val id: String, val title: String, val completed: Boolean)

class GetTasksUseCase @Inject constructor(
    private val repository: TaskRepository
) {
    operator fun invoke(): Flow<List<Task>> = repository.getTasks()
}

Step 3: ViewModel and State Management

5. MVVM with UiState pattern:

// Sealed interface for exhaustive state handling
sealed interface TaskListUiState {
    data object Loading : TaskListUiState
    data class Success(val tasks: List<Task>) : TaskListUiState
    data class Error(val message: String) : TaskListUiState
}

@HiltViewModel
class TaskListViewModel @Inject constructor(
    private val getTasks: GetTasksUseCase
) : ViewModel() {

    val uiState: StateFlow<TaskListUiState> = getTasks()
        .map<List<Task>, TaskListUiState> { TaskListUiState.Success(it) }
        .catch { emit(TaskListUiState.Error(it.message ?: "Unknown error")) }
        .stateIn(
            scope = viewModelScope,
            started = SharingStarted.WhileSubscribed(5_000),
            initialValue = TaskListUiState.Loading
        )
}

6. MVI for complex screens (multiple user actions, side effects):

// Events from UI → ViewModel
sealed interface TaskListEvent {
    data class ToggleTask(val taskId: String) : TaskListEvent
    data class DeleteTask(val taskId: String) : TaskListEvent
    data object RefreshTasks : TaskListEvent
}

// One-time effects (navigation, snackbar)
sealed interface TaskListEffect {
    data class ShowSnackbar(val message: String) : TaskListEffect
    data class NavigateToDetail(val taskId: String) : TaskListEffect
}

Step 4: Dependency Injection with Hilt

7. Module setup:

// In :core:data
@Module
@InstallIn(SingletonComponent::class)
abstract class DataModule {
    @Binds
    abstract fun bindTaskRepository(impl: TaskRepositoryImpl): TaskRepository
}

// In :core:network
@Module
@InstallIn(SingletonComponent::class)
object NetworkModule {
    @Provides
    @Singleton
    fun provideRetrofit(): Retrofit = Retrofit.Builder()
        .baseUrl(BuildConfig.API_BASE_URL)
        .addConverterFactory(Json.asConverterFactory("application/json".toMediaType()))
        .build()
}

8. Hilt rules:
   • Use @Inject constructor (not field injection)
   • Scope to the narrowest lifecycle: @Singleton only for truly app-wide instances
   • Use @Binds for interface→implementation, @Provides for third-party classes
   • Test with @UninstallModules + @BindValue or Hilt test rules

Step 5: Coroutines and Flow

9. Structured concurrency:

// Use viewModelScope — never GlobalScope
viewModelScope.launch {
    repository.syncTasks()
}

// Use withContext for thread switching
suspend fun readFromDisk(): Data = withContext(Dispatchers.IO) {
    file.readText().parseData()
}

// Expose Flow from repositories, collect in ViewModels
fun getTasks(): Flow<List<Task>> = taskDao.observeAll()
    .map { entities -> entities.map { it.toDomain() } }

10. Error handling:

// Catch in the ViewModel, not in the repository
viewModelScope.launch {
    try {
        repository.syncTasks()
    } catch (e: IOException) {
        _uiState.update { it.copy(error = "Network unavailable") }
    }
}

Common Rationalizations

Shortcut	Why It Fails
"We don't need modules, it's a small app"	Modularization pays off at ~10 screens. Build times and code coupling grow fast.
"Use cases are unnecessary boilerplate"	Use cases make ViewModels testable without mocking repositories. They also compose.
"I'll put the logic in the ViewModel"	Fat ViewModels become untestable. Domain logic belongs in use cases.
"Hilt is too complex, I'll use manual DI"	Manual DI doesn't survive multi-module projects. Hilt scales; manual DI doesn't.
"GlobalScope is fine for fire-and-forget"	GlobalScope survives activity destruction, leaking resources. Use structured concurrency.

Red Flags

• Feature module depends on another feature module
• Domain layer imports Android framework classes
• ViewModel directly accesses Room DAOs or Retrofit services
• GlobalScope.launch anywhere in production code
• Field injection (@Inject lateinit var) instead of constructor injection
• Mutable state exposed from ViewModel (MutableStateFlow made public)
• Business logic in Composables

Verification

• Module dependency graph follows layer rules (no upward dependencies)
• Domain layer has zero Android dependencies
• ViewModels expose StateFlow, not MutableStateFlow or LiveData
• All dependencies provided via Hilt constructor injection
• No GlobalScope usage in production code
• Repositories return Flow for observable data
• Use cases are independently testable
• ./gradlew build succeeds with no circular dependency errors