Performance & Caching¶
Measured times¶
Measured on a MacBook Air (Apple M3) with the two samples, running
:app:storeExport :app:storeValidate :app:storeVerify:
| Scenario | KotlinConf (50 assets) | Now in Android (58 assets) |
|---|---|---|
| Cold run (no prior caches) | ~71 s | ~71 s |
clean, restored from build cache |
~6 s | ~3 s |
Rendering costs about 0.3 s per image with Layoutlib, and about 3 s per image for screens that need the Robolectric fallback. Most of the clean-build time is compiling the app itself, which Gradle would do anyway.
What makes it fast¶
- Incremental tasks. Each step declares its inputs and outputs; unchanged steps are skipped.
- Build cache.
storeRenderis cacheable, so CI runs and branch switches reuse renders. - Configuration cache. Supported by every task.
- Shared Layoutlib. Downloaded once and unpacked by an artifact transform into the Gradle user home, shared by all modules and projects.
- Batching. Each renderer JVM renders up to
batchSizeimages, amortizing startup. - Build-wide process limit.
maxParallelProcessesis shared across modules, so parallel builds don't exhaust memory. - Parallel export. Image encoding and hashing run in parallel.
Tuning¶
| Situation | Change |
|---|---|
| Many images and plenty of RAM | Raise rendering.maxParallelProcesses to 3–4. |
| CI runner with little memory | maxParallelProcesses = 1, smaller batchSize. |
| Out-of-memory in renderers | Raise rendererJvmArgs to -Xmx3g. |
| Slow fallback renders | Find the screens with the ROBOLECTRIC badge and make them render in Layoutlib (usually by passing sample state). |
| Only working on one feature | Run the task on that module: ./gradlew :feature:foryou:storeExport. |
Enable the build cache in gradle.properties to benefit from cached renders: