Airbnb’s sequence recommender had a blind spot. JourneyFormer, a Transformer encoder that represents a guest’s bookings, reviews and browsing, ran as a daily batch job, so activity after the run could sit missing from the embedding for nearly two days.
The fix went into Chronon, Airbnb’s open source feature platform. Push Mode makes a streaming job publish a notification when a feature value is committed, so downstream processing reacts to guest activity instead of waiting for a schedule. Near Real Time Model Transform then runs an already deployed model inside that streaming pipeline.
Airbnb reused the same Transformer encoder for the real-time path rather than building a separate inference implementation. Events like searches and listing views combine with short and long-term sequence data, inference fires when the updated state lands, and the resulting embedding writes to the same low-latency store the batch system already used. The retrieval side of ranking stayed unchanged.
The results are modest but measurable. Embedding staleness fell from close to two days to under a minute, with typical updates arriving in 10 to 30 seconds. Offline NDCG improved 1.67 percent, and online A/B testing showed roughly 0.33 percent more uncancelled bookings.
Both capabilities are contributed to the public Chronon repository, where they join its existing emphasis on keeping offline and online computation consistent.
