Building Meta's Internal Cloud: the opportunities and challenges
summary
Meta runs an internal cloud that serves 3.7 billion people on millions of machines and handles Tau’s quadrillion queries per day, using systems like twine for compute and aggregated maintenance trains to reduce operational churn. The team is building a single-pane platform (ICSP) with unified APIs and CLIs to manage code, capacity, data and UX; 21 data center regions and GPU-first ML inference (2023 rollout) aim to scale thousands of models with multi-tenancy and safe gradual rollouts.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:47Meta’s planetary-scale cloud
- 1:49Private to shared compute pools
- 3:58Cross-team collaboration innovations
- 4:40AI and metaverse complexity
- 6:24InfraCloud vision and goals
- 8:102010 infrastructure culture
- 10:59Composable infrastructure needs
- 11:40Cultural shifts for innovation
- 13:03InfraCloud user experience
- 14:51ICSP unified APIs
- 15:55ICSP impact on ML inference
- 17:44Journey in culture and tech
- 18:26Summary
related talks
21mExtending Meta to the Public Cloud
23mSupporting AI holistically Across the Infra stack
21mCapacity Planning for AI Training Hardware
23mKeynote
50mWhen Should You Consider MetaArchitectures? Using Meta to Scale the Cloud
44mBuilding Multi-Tenant Compute Systems in the Enterprise
15mHybrid Cloud Ecology of Tencent, Reconstructing the future digital infrastructure
29mLive Keynote