Optimizing Billion‑Scale Video Playback: Architecture, Bandwidth, Startup, Buffering, and Success‑Rate Improvements

Rapid Deployment

The team, responsible for both mobile QQ and QQ Space on iOS and Android, had to implement four parallel codebases, which they streamlined by introducing a local proxy that mediates between the player and the server, enabling unified download control, pre‑loading, priority scheduling, and detailed monitoring.

Key technical points include edge‑streaming (download‑while‑play), traffic shaping, code reuse across the four teams, support for third‑party video sources, and a robust reporting system that captures every error code and playback trace.

Cost Optimization

After a month of launch, playback rose fourfold while bandwidth surged sixfold; to curb waste, the team applied multi‑level rate limiting, pre‑download windows, and introduced H.265 encoding for hot videos, achieving roughly a 31% bandwidth reduction.

They evaluated device decoding capabilities, performed large‑scale floating‑point benchmarks to decide H.265 suitability, and limited transcoding to top‑N videos, which account for 80‑90% of traffic.

Startup (秒开) Optimization

The goal was sub‑second startup, matching Facebook’s instant‑play experience. Strategies involved probing MP4 containers to locate the MOOV atom, using a download proxy to fetch only essential header data, and employing parallel HLS segment downloads to cut startup from 6 seconds to about 1.6 seconds.

Additional measures included adaptive bitrate throttling, pre‑loading of key frames, and a custom file‑hole cache that reduces seek‑induced buffering from 4.6% to 0.48%.

Success‑Rate and Reliability

Comprehensive error‑code instrumentation and real‑time alerting raised download success from 97.1% to 99.9% and playback success from 97.0% to 99.9%; first‑buffer latency dropped from 1.95 s to 0.7 s, and second‑buffer probability fell from 4.63% to 0.48%.

Q&A

During the Q&A, the speaker confirmed that roughly 80% of hot videos now use H.265 and explained the rationale behind using large‑scale floating‑point benchmarks to assess device decoding capability.