NEWS · MODELS · #2726
Airbnb adds Push Mode and NRT Model Transform to Chronon to update embeddings in seconds
Airbnb extended its open-source feature platform Chronon with Push Mode and Near Real Time (NRT) Model Transform so the deployed Transformer-based JourneyFormer encoder can run inside a streaming pipeline. The change reduced sequence-embedding staleness from nearly two days to under a minute (typical 10–30s) and Airbnb reports a +1.67% offline NDCG and ~0.33% increase in uncancelled bookings in online A/B tests; the features were contributed to the public Chronon repository.
KEY POINTS
- Airbnb extended its open-source feature platform Chronon with Push Mode and Near Real Time (NRT) Model Transform so the deployed Transformer-based JourneyFormer encoder can run inside a streaming pipeline.
- The change reduced sequence-embedding staleness from nearly two days to under a minute (typical 10–30s) and Airbnb reports a +1.67% offline NDCG and ~0.33% increase in uncancelled bookings in online A/B tests; the features were contributed to the public Chronon repository.
- Bringing model inference into Chronon's streaming path lets Airbnb keep online representations much fresher with minimal serving changes, improving personalization metrics and providing an open-source pattern for near‑real‑time feature-driven inference.
WHY IT MATTERS
Bringing model inference into Chronon's streaming path lets Airbnb keep online representations much fresher with minimal serving changes, improving personalization metrics and providing an open-source pattern for near‑real‑time feature-driven inference.