Low‑Latency Video Streaming Optimizations for Douyin During the World Cup

Background

The World Cup live stream on Douyin achieved 4K ultra‑HD with a peak concurrent PCU of over 37 million, posing a massive low‑latency challenge. The article first outlines the overall signal distribution chain and the need to accurately measure latency at each production and transmission stage.

Full‑Chain Latency Measurement

Production latency is largely supplier‑controlled; the technical team focuses on precise measurement and supplier coordination. Transmission latency is more controllable and becomes the primary optimization target, aiming to reduce delay without degrading playback experience.

Production Stage Optimizations

Signal sources involve multiple upstream steps with limited influence. Studio production adds ~1.5 s delay due to commentary and packaging, which was accepted to maintain stability.

Transmission Stage Optimizations

Key latency contributors are transcoding, CDN distribution, and playback buffering. Real‑time transcoding adds <300 ms, while CDN adds modest delay. Buffer‑induced latency can exceed 5 s, so strategies focus on minimizing buffering while preserving QoE.

FLV Solution

FLV is the primary low‑latency protocol; a 2‑second FLV solution was insufficient for the World Cup, prompting a 2‑second FLV with frame‑catch‑up and drop‑frame tactics. Detailed buffer‑and‑stall‑based dual‑threshold speed‑up logic uses current buffer length, variance, and stall history to decide safe speed‑up, limiting acceleration to ≤2 s and avoiding frequent speed changes.

RTM Solution

RTM, based on WebRTC, targets sub‑1‑second end‑to‑end latency. Initial AB tests showed lower latency but poorer pull‑stream success, first‑frame time, and stability compared to FLV. Optimizations included DNS node selection, MiniSDP binary signalling over UDP, pre‑loading signalling, UDP node probing, jitter‑buffer adjustments, and custom playback control to reduce stalls.

MiniSDP Signalling

MiniSDP compresses SDP to ~300 bytes sent in a single UDP packet, dramatically improving signalling success rate and first‑frame latency.

Pull‑Stream Success Rate

Network‑grade based user selection routes high‑quality users to RTM, while others fall back to FLV, improving overall success.

Stall Reduction

Disabling aggressive frame dropping in the RTC kernel and applying fine‑grained speed‑up policies reduced stalls, aligning RTM QoE with FLV.

World Cup Performance Comparison

Douyin consistently maintained ~30 s lower latency than competing products, and aggressive catch‑up strategies that reduced latency quickly were found to degrade user experience, so they were not adopted.

Future Directions

Continued work will refine FLV and RTM, improve pull‑stream success, extend RTP features, and standardize RTM across CDNs. Exploration of CMAF/LL‑HLS with QUIC and connection reuse aims to lower chunk‑based latency. XR live streaming will adopt tile‑based transmission with priority‑aware UDP to handle 8K‑level streams while keeping latency and bandwidth low.