Problem
LINE VOOM is a video social network inside the LINE app. Its "For you" tab drew recommendations from a candidate pool built once a day by an offline batch job that covered embedding generation, storage and similar-post search. A new post stayed out of recommendations for up to a day, so fresh content missed the moment when it mattered. Fixing this meant storing post embeddings online and searching for similar posts at request time, which required a vector database at scale.
Example from the article: A creator posts a "Happy New Year" video, but it is not recommended to users until the next daily candidate update.
Solution in brief
Replace the daily batch with model-based candidate generation: embed each new post, write it to an online vector database, and run approximate nearest neighbor (ANN) search per request. LINE chose Milvus in cluster mode and verified it with chaos tests and performance tests on its own infrastructure. An HNSW index, high-availability (HA) setups and a daily backup made it fast and stable enough.
Questions this case answers
- How to make newly created items eligible as recommendation candidates immediately instead of after a daily offline batch
- How to choose and validate a self-hosted vector index service for low-latency retrieval when public cluster-scale use cases are few
- How to make a fresh video post a recommendation candidate within the same day instead of after a daily batch?
- How to choose a self-hosted vector database for low-latency candidate retrieval?
Tools and methods
- Milvus (cluster) — online vector storage and similarity search
- HNSW index — ANN index chosen for low latency and high throughput
- Chaos test — pod kills and scaling while API requests keep flowing
- Collection alias with dual writing — switches search to a backup collection
- Milvus backup server, MinIO, Airflow — daily collection backup
- Birdwatcher — Milvus tool that reads component state from etcd
Google or a one-time email code. No password.
