Compose performance

Compose Performance — Stability, Recomposition, Metrics

Instructions

Compose is fast when it can skip. Most performance complaints are stability problems in disguise. Measure first; fix the root cause.

1. Generate Compose Compiler Reports

Add to each Compose module's build.gradle.kts:

tasks.withType<org.jetbrains.kotlin.gradle.tasks.KotlinCompile>().configureEach {
    compilerOptions.freeCompilerArgs.addAll(
        "-P", "plugin:androidx.compose.compiler.plugins.kotlin:reportsDestination=${layout.buildDirectory.get()}/compose-reports",
        "-P", "plugin:androidx.compose.compiler.plugins.kotlin:metricsDestination=${layout.buildDirectory.get()}/compose-metrics",
    )
}

Build and inspect build/compose-reports/*-classes.txt. You want:

• Every data class used as Compose state: stable.
• Every screen composable: skippable and restartable.
• Zero unstable parameter: lines in hot paths.

2. Common Instability Causes and Fixes

Cause	Fix
List<T> parameter	Use ImmutableList<T> from kotlinx.collections.immutable.
LocalDateTime, LocalDate	Annotate a wrapper @Immutable data class or map to Long.
Types from third-party libs	Wrap in your own @Immutable class.
var field in a data class used as state	Change to val; create a new instance to update.
Function-type parameter (lambda) capturing unstable value	Hoist lambda into remember(key) { { ... } } or accept a stable action/state.

3. Recomposition Counts in Layout Inspector

Android Studio → Layout Inspector → enable "Show recomposition counts" on a running app. Scroll your screen; any composable with growing counts during a static state is re-running too often. Pair with compiler reports to pinpoint the unstable parameter.

4. Common Fixes Cookbook

a) Stable lambda

// BAD — new lambda identity on every recomposition, may defeat skipping downstream
ArticleList(items, onClick = { id -> vm.onAction(Open(id)) })

// GOOD — stable reference
val onClick: (String) -> Unit = remember(vm) { { id -> vm.onAction(Open(id)) } }
ArticleList(items, onClick = onClick)

Do this only where the report shows restartable skippable skipped: NO. Premature lambda hoisting bloats code for no gain.

b) Immutable lists

data class ArticlesUiState(val items: ImmutableList<Article> = persistentListOf())

c) derivedStateOf to avoid reading too much state

val canSubmit by remember { derivedStateOf { form.name.isNotBlank() && form.agree } }

Without derivedStateOf, every form change recomposes the submit button; with it, only when the boolean actually changes.

d) Keyed LazyColumn items

items(items = list, key = { it.id }, contentType = { "article" }) { Article(it) }

Missing key causes "reset" recomposition on list changes — every visible item rebuilds.

e) Defer reads with lambdas

// BAD — text recomposes whole parent
Text("Count: ${state.count}")

// GOOD — only Text recomposes
Text(text = { "Count: ${state.count}" }.toString()) // when accepting the value
// Or, for graphicsLayer, pass a lambda that reads state inside the block:
Modifier.graphicsLayer { alpha = state.alpha }

Prefer Modifier.graphicsLayer { ... } (block form) over Modifier.alpha(state.alpha) for animated values — the state read moves to draw phase, skipping composition and layout.

5. Baseline Profiles

Generate a Baseline Profile to speed up critical-path composables AOT:

// baseline-profile module
@OptIn(ExperimentalBaselineProfilesApi::class)
class BaselineProfileGenerator {
    @get:Rule val rule = BaselineProfileRule()

    @Test
    fun startup() = rule.collect(packageName = "com.example.app") {
        pressHome(); startActivityAndWait()
        device.findObject(By.res("feed_list")).fling(Direction.DOWN)
    }
}

Run with the Baseline Profile Gradle Plugin. Ship the generated baseline-prof.txt with the app.

6. Rules of Thumb

• 60/120 Hz budget: each frame = 16 ms / 8 ms. Keep composition + layout + draw under that for visible screens.
• Prefer @Composable extraction only when it creates a stable boundary. Tiny helpers that take unstable inputs aren't helping.
• Do not call remember { expensive() } without a key — if inputs change the value is wrong. Use remember(inputs) { }.
• SnapshotStateList (from mutableStateListOf) is great for local UI state but is unstable as a function parameter unless declared explicitly; prefer ImmutableList for cross-boundary.

7. Macrobenchmark

@get:Rule val benchmarkRule = MacrobenchmarkRule()

@Test
fun scrollFeed() = benchmarkRule.measureRepeated(
    packageName = "com.example.app",
    metrics = listOf(FrameTimingMetric()),
    iterations = 10,
    startupMode = StartupMode.WARM,
    compilationMode = CompilationMode.Partial(BaselineProfileMode.Require),
) {
    startActivityAndWait()
    device.findObject(By.res("feed_list")).fling(Direction.DOWN)
}

Run in CI for every release and treat frameDurationCpuMs P95 > 16 ms as a regression.

Checklist

• Compose compiler reports generated for every Compose module.
• No screen-level composable has restartable skippable skipped: NO in the report.
• Lists use ImmutableList / persistentListOf().
• LazyColumn/LazyRow items all pass key.
• Animated values use Modifier.graphicsLayer { } block form.
• A Baseline Profile is generated and bundled with the release build.
• Macrobenchmark frame-timing tests run in CI.