TikTok's For You Page is the model recommendation system of 2020s. Watch time + engagement → immediate feedback → next-video ranked. Global scale.

UserSignalswatch, like, shareFeed APIEvent IngestKafkaCandidate GenrecallFine RankDNNRe-rank + diversifyUser Feature StoreReal-time model updateVideo CDNEmbedding StoreOnline Feature1min freshnessModel Store
TikTok FYP: candidate gen + rank + real-time signals → next video
Advertisement

Short feedback loop

Watch time per video = instant signal. Skip = negative. Like/share = positive. Model updates in near-real-time.

Short feedback loop

Watch time per video = instant signal. Skip = negative. Like/share = positive. Model updates in near-real-time.

Advertisement

Real-time feature store

User's recent interactions available within 1 minute for next request. Fresh signals matter for short videos.

Candidate → fine → rerank

Candidate gen recalls 1000s. Fine ranker (DNN) scores. Rerank ensures diversity + prevents repetition.

Model updates continuously

Not batch retrain. Streaming updates to embeddings + parameters. Emerging trends captured within hours.

A/B at massive scale

Thousands of experiments running simultaneously. Small user cohorts test model variants. Winners rolled out.

Real-time signals + fresh features + continuous model updates. Feedback loop as tight as any consumer product.