New Performance Challenges in the Agentic Era for Android Builds

on YouTube ↗44m

" summary

Measured data shows 3,300 local Gradle builds in one week with 1,700 agentic builds (52%), and a VS Code Kotlin extension that spawns a Gradle call per module can trigger hundreds or thousands of builds just by opening a project. Parallel agents share a single Kotlin daemon, creating GC contention — fix by increasing Kotlin memory or using in-process compilation. Experiments with nine parallel builds show a remote cache (Radius) is ~10% faster overall with a 36% higher cache-hit rate, and CI caching combos cut a 449s clean build to ~200s or 69s with local cache.

AI-written summary · about

" chapters (13)

1. 0:00Agentic builds introduction
2. 1:26AI-driven local builds
3. 3:38Lighter editors and builds
4. 8:50Parallel builds with Git work trees
5. 12:26Parallel builds and caching
6. 17:01CI builds surge with AI
7. 19:30Dependency graphs reduce CI
8. 21:19LLMs increase test workload
9. 24:12CI caching strategies
10. 32:41Specialized AI models for builds
11. 38:47ML for CI agent allocation
12. 40:10Gradle 9.5 AI features
13. 42:42Summary and future

" related talks

• 30mImproving Android Build performanceTony Robalik · droidCon NYC 2018
• 31mImproving Android Build performanceTony Robalik · droidcon San Francisco 2018
• 44mGradle: The Build System That Loves To Hate YouAurimas Liutikas · droidcon London 2024
• 27mFaster builds on Android: from 2 minutes to 2 secondsPhilipp Hofmann · droidcon Italy 2019
• 33mIMPROVING ANDROID BUILD PERFORMANCEStefan Oehme · droidcon Berlin 2018
• 28mMaking the most of your Gradle BuildsEgor Andreevici · Droidcon Italy 2017
• 48mOptimizing Android Build PerformanceHans Dockter, César Puerta · droidcon San Francisco 2017
• 30mHow we brought down our build times. General optimization tipsManuel Nakamurakare · droidcon San Francisco 2022