Airbnb has changed its recommendation infrastructure so recent searches and listing views can influence a guest’s recommendations within seconds instead of waiting for a daily batch. The company extended its open-source Chronon feature platform with Push Mode and Near Real-Time Model Transform.
Previously, Airbnb’s JourneyFormer sequence model encoded booking, review and browsing history once per day. Processing delays meant the stored representation could be almost two days out of date. The new pipeline sends a notification when feature data is committed, combines the event with short- and long-term history, and runs the already deployed model inside the stream. The refreshed embedding is written to the same low-latency store used by the existing ranking system.
That design leaves the latency-sensitive search path largely unchanged while making its retrieved user representation fresher. Airbnb reports typical updates in 10 to 30 seconds, a 1.67% improvement in offline normalized discounted cumulative gain, and about a 0.33% lift in uncancelled bookings during an online test. Those figures come from Airbnb’s own deployment and may not transfer directly to other recommendation systems. Push Mode and the model-transform capability are available in Chronon’s public repository.