Airbnb Reduces Recommendation Staleness from Days to Seconds Using Chronon Updates

Real-Time Recommendations Implemented

Airbnb has enhanced its open-source Chronon feature platform by integrating Push Mode and Near Real Time (NRT) Model Transform. This development allows sequence-based recommendation embeddings to update within seconds of guest activity. Previously, these embeddings could be up to two days old, but now they typically update within 10 to 30 seconds.

Impact on User Experience and Metrics

The reduction in data staleness has shown measurable improvements. Airbnb reported a 1.67% improvement in offline Normalized Discounted Cumulative Gain (NDCG), a common metric for ranking quality. Online A/B testing also indicated an approximate 0.33% increase in uncancelled bookings, suggesting more relevant recommendations lead to better user engagement and conversion.

Addressing Previous Latency Issues

The previous system relied on a daily batch job to run a Transformer encoder, which processed guest booking, review, and browsing history to generate embeddings. While these embeddings were stored for low-latency retrieval during searches, any guest activity occurring after the daily batch would not be reflected until the next day's run, leading to significant delays in recommendation freshness.

Chronon's New Capabilities

To resolve the staleness problem, Airbnb introduced two new features to Chronon. Push Mode enables streaming jobs to publish notifications when feature values are committed, allowing immediate downstream processing. NRT Model Transform allows deployed models, such as the Transformer sequence encoder, to run directly within this streaming pipeline. This avoids the need for a separate inference implementation for streaming workflows.

Updated Recommendation Pipeline

The new pipeline captures real-time events like searches and listing views, combines them with existing guest sequence data, and triggers model inference when updated feature states are available. The newly generated embeddings are then written to the same low-latency store used by the batch system. This integration means the online ranking system continues to function as before, but with significantly more current user representations.