Rajomon: Decentralized and Coordinated Overload Control for Latency-Sensitive...
summary
RAJOMON is a market-style, decentralized overload controller where clients attach tokens and prices to requests and services charge per-interface prices, dropping requests with insufficient tokens; prices and tokens piggyback through the call graph for end-to-end coordination. Implemented as a lightweight gRPC/Go library and evaluated on setups up to 140 cores, it claims up to 78% tail-latency reduction and 45% goodput increase versus prior overload controllers, and it both rate-limits clients and sheds load in-place.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 0:43Microservices and overload
- 2:54Challenges in load control
- 3:57Microservice characteristics
- 5:41Design requirements
- 6:55Existing solutions
- 8:08The Rajomon design
- 10:00Mechanisms and policies
- 12:18Evaluation and experiments
- 14:34Conclusion
related talks
47mMonolith to Microserv. at 20M+ requests per second: Latency and Scale Challenges
22mA Case Study of Enhancing gRPC Micro-Services for Supporting High Throughput, Low Latency Trading
58mRSocket – new Reactive cross-network Protocol
44mDeep Dive talk | Resilience in Microservices with Congestion Control
42mStop Rate Limiting! Capacity Management Done Right
13mLow-Maintenance Backend Architectures for Scalable Applications
17msub0 Asia 2024 - Jakub Panik (GalacticCouncil)
22mAsynchronous Computing at Scale: Overview and Learning