Technical Case Study: Accelerating Live‑to‑VOD Conversion for the 2025 Spring Festival Gala

Project Background

The 2025 Spring Festival Gala is a large‑scale live broadcast. After the live stream ends, the time required to generate on‑demand (VOD) assets is a critical competitive metric. Competitors can produce VOD from more than 4 hours of live content within 10 minutes. The existing rapid direct‑to‑VOD system required at least 40 minutes for the same workload, far from the business goal.

By redesigning the entire live‑to‑VOD pipeline, the new system produced VOD for a 4 h 40 min gala in about 8 minutes, achieving roughly a 5× speedup.

Problem Analysis – Longest Path

The longest path in the original workflow consists of three stages:

Final VOD transcoding

Operations backend clipping

VOD asset production

Each stage was identified as a major bottleneck. For example, VOD transcoding runs synchronously with the live stream and uses large segment sizes; shortening segment length yields the most direct speed gain. The operations backend required manual clipping of long recordings, which was time‑consuming due to a heavy legacy UI.

Migration from FLV to HLSv7

The legacy system stored live recordings as FLV files with 3‑minute segments. FLV is not a segment‑friendly format, and converting FLV to DASH for VOD introduced heavy I/O and storage overhead. Switching to HLSv7 (using fMP4 fragments) aligns with modern streaming protocols, enables second‑level segmenting, and allows browsers to load only an m3u8 playlist instead of the whole file.

Technical Solution

New Live Recording System : Replaces the old recorder with an m3u8‑based service that reports segment information in real time, dramatically improving QPS and supporting fine‑grained (second‑level) segment updates.

New Direct‑to‑VOD Transcoder : A state‑machine‑driven, short‑segment, high‑throughput transcoder built on the next‑generation media processing platform. It supports sub‑second segment processing and outputs directly to DASH/MP4, bypassing the costly FLV‑to‑DASH conversion.

System Work Model : Designed for maximum concurrency through three mechanisms – horizontal compute parallelism, local I/O + compute overlap, and multi‑stage pipelining. Segment length can be tuned per resource availability.

Event‑Driven m3u8 Proxy : Utilizes OpenBayes Ray Actor to create a lightweight distributed queue. The proxy performs full‑global m3u8 updates locally, eliminating blind reads from distributed storage and reducing storage pressure. New segments are pushed via messages, and a local m3u8 copy serves high‑frequency reads.

Audio‑Video Sync Assurance : Timestamp cleaning, packet re‑assembly, and re‑segmentation guarantee AV sync even with thousands of tiny segments. The solution has been validated in both the new and legacy pipelines.

New Backend UI : The updated UI removes the need for manual loading of large FLV files. It leverages hls.js for instant playlist loading, provides timeline controls, and enables “instant clipping” with sub‑second precision.

Production Guarantees

Two‑phase assurance strategy:

Development phase – built standardized playback gateway and automated AV‑sync detection to surface issues early.

Execution phase – deployed three‑fold live stream backup, edge‑to‑center recording redundancy, OpenBayes dedicated cluster reset before the event, and comprehensive monitoring/alerting. Results The optimized pipeline reduced the total processing time from an estimated 41 minutes (old system) to about 8 minutes (new system). Specific improvements include: Transcoding: from 6 minutes to ~30 seconds. Operations backend: from 5 minutes to <10 seconds. Virtual original‑clip synthesis: from 10 minutes to 0 seconds. FLV‑based VOD generation: from 10 minutes to 7 seconds (DASH/MP4). Overall, the project achieved a ~5× acceleration, validated the next‑generation streaming infrastructure, and laid the foundation for future large‑scale live‑to‑VOD use cases.