The guest journey, updated in real time: extending Airbnb’s sequence recommender with Chronon

Airbnb boosted offline search ranking quality by +1.67% NDCG and uncancelled bookings by a third of a percent on a decade-old ranking system—not by changing the core model, but by refreshing guest embeddings in 10 to 30 seconds instead of two days. 👇

Airbnb previously computed guest search embeddings using a nightly batch job. When a user ran a search, the ranking model fetched their sequence embedding from a low-latency store. But because the batch pipeline only ran once a day and suffered from processing lag, embeddings were often up to two days stale. A guest actively browsing stays on a Tuesday afternoon received search results that completely ignored what they had clicked on 10 minutes prior.

To fix this without running heavy multi-layer Transformer inference on the critical serving path, Airbnb added two general-purpose capabilities to Chronon, its open-source feature platform: Push Mode and Near-real-time Model Transform.

The entire update process now runs asynchronously in the background as a continuous event flow:

Guest event → Stream join → Push trigger → Streaming inference → Feature store

First, when a guest views a listing or runs a search, a streaming job captures the event and merges it with the guest's long-term and short-term sequence history.

Push Mode then fires a lightweight notification the instant this feature state updates, turning state changes into event-driven triggers rather than waiting for scheduled jobs to poll for changes.

Near-real-time Model Transform receives this notification and calls the Transformer sequence encoder directly from within the streaming pipeline, writing the resulting embedding straight back to the low-latency store as a feature value.

When the guest submits their next search, the online ranking model retrieves this fresh embedding from the store using the exact same serving path as before.

By shifting from daily batch jobs to event-driven streaming inference, Airbnb cut feature staleness from roughly 48 hours down to 10–30 seconds.

Offline evaluations showed a +1.67% jump in NDCG over the daily-batch baseline. Online A/B testing confirmed an approximate 0.33% increase in uncancelled bookings, proving that feature freshness alone drives significant ranking quality gains.