Xiaohongshu handled 50x World Cup traffic: +50% bitrate, -30% latency, 57x faster 4K transcoding
机器之心 · wechat · 2026-10-21
Xiaohongshu (RedNote) detailed the engineering behind streaming the 2026 World Cup as an official rights holder, where viewership hit 50x its normal live traffic and interactions 100x. Results: 50% higher live consumption bitrate, 30% lower end-to-end latency, and 57x faster 4K enhanced replay transcoding.
Quality & latency
- Super-resolution fused with artifact repair; temporal causal models reduce detail flicker; content-aware strength on grass suppresses breathing effects; frames split into tiles across multiple GPUs to cut SR latency.
- Frame interpolation fine-tuned on soccer footage, processing only arrived frames to avoid buffering.
- NTP clock alignment plus SEI time anchors measure per-stage delay; encoder latency cut from 1.2s to under 400ms.
- HTTP-FLV main path with RTC low-latency for eligible users; player jointly tunes startup watermark, buffering, and catch-up.
3-week delivery at 10M+ concurrency
- Componentized client (RN for complex panels, lightweight DSL for first-frame-sensitive widgets); preloading off-peak with randomized request spreading; layered fallbacks (dual degradation, SafeMode, IM isolation) rehearsed pre-match and switched via config.
- Long-session profiling found the interaction layer drove thermal throttling; value-tiered degradation preserved core playback first.
4K replays
- Offline SR/interpolation/color models; multi-GPU per node (+2.3x) plus higher distributed shard limits.
- Logical sharding preserves Open-GOP decode context; VFR normalized to CFR with 5s keyframes for 100% segment alignment across ABR ladders.
- "Fast-then-refined": quick ladders first, then high-compression re-encode before peak viewing, cutting 4K bitrate another 13.3% at equal quality.
The post ends with recruiting info for the multimedia team (2027 grads and experienced engineers).
More from Infra
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Jensen Huang lays out NVIDIA's full Vera Rubin agentic system: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
- ONNX to Transformers.js: running models directly in your browser — JosephJacks_ · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- "We bet on Cloudflare and it paid off": R2 and D1 at billion scale — dinasaur_404 · 2026-10-23
- Public data is drying up; private training data is turning into a browsable catalog — SucceededMind · 2026-10-23