Production Multi-Party Computation via the Distributed Aggregation Protocol
summary
A production deployment called Divvi Up runs the IETF Distributed Aggregation Protocol to process billions of secret-shared telemetry reports over 3+ years, routinely bursting past 450 million contributions per week. Clients split measurements into two shares sent to independent leader and helper servers; verifiable aggregation via VDAF rejects malformed inputs and reconstructs only sums, counts, and histograms. The aggregator is a Rust web service on EC2 with managed Postgres, interoperable client libs (TypeScript, Rust, Java), and composes with differential privacy and OHTTP.
AI-written summary · about
chapters (14)
- 0:00Multi-party computation at scale
- 1:35Distributed Aggregation Protocol
- 3:21Two servers balance privacy
- 5:27VDAFs enable high-volume telemetry
- 6:29Independent server challenge with Divvy Up
- 6:51Production proof with billions reports
- 7:55Firefox telemetry privacy
- 9:19Cost breakdown for DAP
- 9:41Privacy with differential privacy
- 10:22Metadata privacy with oblivious HTTP
- 11:45Start using DAP today
- 13:48Lowering MPC barriers
- 14:09Resources for DAP implementation
- 14:52Q&A
related talks
01h 07mIndustry Presentations
19mDisco: Efficient Distributed Window Aggregation
32mPrivate aggregation at scale with decentralized trust
29mBuilding Probabilistic BFT Protocols Through Subsampled Voting
14mData-Parallel Actors: A Programming Model for Scalable Query Serving Systems
48mDistributed Consensus Algorithms
45mActors Can Rule Your DDD World
22mPARTISAN: Scaling the Distributed Actor Runtime