Airbnb Chronon platform updates sequence embeddings to seconds, reducing staleness from two days

Updates to Airbnb's Chronon feature platform, including the addition of Push Mode and Near Real Time (NRT) Model Transform, now allow sequence-based recommendation embeddings to be refreshed within seconds following guest interactions. This significant architectural enhancement has dramatically cut the data staleness for embeddings, moving it from nearly two days down to less than a minute, with typical updates arriving in a span of 10 to 30 seconds. Furthermore, the company observed a 1.67% boost in offline NDCG metrics and saw an approximate 0.33% lift in uncancelled bookings during online A/B testing.

This progress builds upon Airbnb's existing JourneyFormer sequence recommender, which leverages a Transformer encoder to model a user's history encompassing browsing, reviewing, and booking activities. Previously, the encoder operated as a daily batch process, generating embeddings that were then housed in a low-latency data repository. While the ranking system would fetch the most recent embedding during a search and merge it with the live query, keeping the Transformer outside the critical, latency-sensitive serving path, any activity that occurred after the daily batch cycle could remain unrepresented in the embedding until the subsequent run, leading to effective staleness approaching two days.

To resolve this freshness issue, Airbnb integrated two new functionalities into Chronon. Push Mode initiates a streaming job to issue a notification whenever a feature value is finalized, enabling subsequent processing to react to updated guest behavior instead of relying on scheduled checks or polling. Concurrently, NRT Model Transform permits an already deployed model to function directly within that streaming data flow. Airbnb utilizes the same Transformer sequence encoder for this near real-time path, thereby circumventing the need for a separate inference implementation for the streaming workflow.

The resulting pipeline captures events such as listing views and searches, integrates them with the guest's established short-term and long-term sequence data, and triggers model inference when the updated feature state becomes available. This newly generated embedding is then persisted into the identical low-latency store utilized by the batch system, meaning the online ranking path itself experiences minimal alteration, yet the data it accesses is substantially more current.

Pengyu Hou, a Senior Software Engineer focused on Infrastructure at Airbnb, characterized this platform evolution as bridging the divide between data generated by events and the execution of model inference. He noted that "NRT Model Transform lets Chronon call an already deployed model ... directly from within the streaming flow." This methodology aligns with a broader feature serving paradigm where predictions can be refreshed as underlying features change, rather than being recalculated for every single user request.

While other open-source feature platforms, such as Feast, document cached prediction architectures where inference is triggered by feature updates, thereby exchanging increased pipeline complexity and extra writes for reduced serving latency, Chronon uniquely embeds streaming computation and model transformation directly into its feature platform, placing a strong emphasis on ensuring consistency between offline and online computations.

Airbnb has made both Push Mode and NRT Model Transform available in the public Chronon repository. This platform is designed to support both batch and streaming computation, offering online feature serving capabilities built to maintain alignment between the training phase and the production inference environment.

Why it matters

This demonstrates a significant shift in feature serving patterns, moving from daily batch processing to event-driven updates for model inference. Developers interested in low-latency systems can examine how Chronon integrates streaming computation and model transformation for consistency.