Airbnb Cuts Recommendation Embedding Latency from Two Days to 30 Seconds

Airbnb has added Push Mode and near-real-time (NRT) model transformation capabilities to its open-source feature platform Chronon, enabling sequence-based recommendation embedding vectors to be updated within seconds of a guest's action. This optimization reduces embedding lag from nearly two days to less than a minute, typically completing updates in 10 to 30 seconds. Airbnb also disclosed that offline NDCG improved by 1.67%, and online A/B tests showed an approximately 0.33% increase in non-cancelled bookings.

The solution is built on Airbnb's JourneyFormer-based sequence recommender, which uses a Transformer encoder to represent a guest's booking, review, and browsing history. Previously, the encoder ran as a daily batch job, and the generated embedding vectors were stored in low-latency storage. During search, the ranking system retrieves the latest embedding vectors and combines them with the real-time query, keeping the Transformer outside the latency-sensitive online inference path. However, activity occurring after the daily batch job might not be reflected in the embedding vectors until the next run, and combined with processing delays, the actual lag was close to two days.

Airbnb solved the data freshness problem by adding two capabilities to Chronon: Push Mode allows streaming processing to proactively send notifications after a feature value is committed, enabling downstream systems to respond to guest behavior changes without polling or waiting for scheduled jobs; near-real-time model transformation allows deployed models to run directly within the streaming pipeline. Airbnb uses the same Transformer sequence encoder in the near-real-time path, avoiding the need to separately implement inference for the stream processing workflow.

The new pipeline collects events such as searches and listing views, merges new events with the guest's existing short-term and long-term sequence data, and triggers model inference once the feature state update is complete. The generated embedding vectors are written to low-latency storage shared with the batch processing system. As a result, the online ranking path remains largely unchanged, but the retrieved representations become more timely.

Airbnb senior software engineer for infrastructure Pengyu Hou described the platform change as bridging the gap between event-driven data and model inference:

The near-real-time model transformation capability allows Chronon to directly invoke already deployed models within the streaming path...

This approach follows a broader feature serving pattern in which predictions can be updated when the underlying features change, rather than being computed on every user request. Another open-source feature platform, Feast, also has a cached prediction architecture: inference is triggered by feature updates, trading higher write volume and pipeline complexity for lower serving latency. Chronon, by contrast, integrates stream computing and model transformation into the feature platform, with a focus on ensuring consistency between offline computation and online inference.

Airbnb has contributed Push Mode and near-real-time model transformation to the public Chronon codebase. Chronon supports both batch and stream computing and provides online feature serving designed to maintain consistency between training and production inference.

View the original English article: https://www.infoq.com/news/2026/10/airbnb-chronon-recommendation/