Realtime performance at scale (...)
summary
Claims extreme throughput: uses Erlang NIF fanout plus batching to push binary payloads and avoid copies, tunes Ranch sockets to active once with large kernel buffers, and splits pub/sub into connection/mux/Q node types so subscriptions are per-subscriber processes. Measures show about 250k msgs/s between vanilla nodes before the NIF, and after the low‑level optimizations and reduced garbage generation the system reaches millions msgs/s with low queuing delay and close to zero contention.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:40Fredriks background
- 2:26Game of War
- 4:22Gameplay
- 6:18More about the game
- 7:58Realtime translations
- 9:40Performance and upscale
- 10:50Map services
- 12:26Map implementation
- 15:57Low latency
- 19:09Performance
- 22:04patch messages
- 27:47data structures
- 30:25multiplexing
- 33:49Connection process
- 34:44Message passing
- 36:19Pulls
related talks
41mPushing RabbitMQ to the limit
33mReaching Ludicrous Speed
39mSuper Fast Real-time Data Processing on Cloud-Native Architecture
26mAre You Sure About Your Mesh Performance?
01h 29mWhen Nanoseconds Matter: Ultrafast Trading Systems in C++
33mConversational Big Data
40mPerf Stories
38mPerformance in the Wild