Asynchronous Computing at Scale: Overview and Learning
summary
Meta reports the platform handles more than a trillion asynchronous workloads per day after re-architecting from a pull-based dispatcher reading Fox (Facebook order queuing service) to a push-based scheduler and lightweight per-source Tailor adapters. They keep submitter + Fox for queues, add stream readers that use synchronous RPCs, and offload retries to a delay service. Result: lower end-to-end latency, more even multi-region load, no data duplication for external sources, and a clear tradeoff: streams + RPCs win for short high-volume jobs while queues keep granular retries.
AI-written summary · about
chapters (9)
- 0:00Start
- 3:20Standard Building Blocks
- 4:08Back in 2020
- 7:11Increasing Complexity of the System
- 10:44Pull Mode. Complex Integration
- 11:35Pull Mode. Spikes in Latency
- 12:41Pull Mode. Load Balancing
- 18:49Final Architecture
- 20:19Learnings
related talks
01h 02mAsynchronous Distributed Systems
49mNeed for ASYNC: in pursuit of scalable internet-scale applications
01h 01mThe Unsung Hero of Modern Software: Asynchronous Messaging
40mEverything Will Flow
44mFrom Synchronous to Asynchronous Distributed Microservices
53mYou Keep Using That Word: Asynchronous And Interprocess Comms
42mIt's like 10,000 streams when what you need is a queue
57mManaging Asynchronous APIs