KOEN
Read Original
#AI

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

The guest journey, updated in real time: extending Airbnb’s sequence recommender with Chronon
01

Summary

How Airbnb Cut Search Recommendation Latency from 2 Days to 10 Seconds

Overcoming the real-time ML bottleneck by extending the Chronon platform with Push Mode and Near-real-time Model Transform

Airbnb has overhauled its personalization model by integrating real-time guest interaction data into its search ranking using Chronon. By transitioning from scheduled daily batch runs to event-driven streaming pipelines, the platform now updates user search embeddings in near-real-time. This article details how they successfully balanced system latency, offline-online consistency, and model performance.

  • 01Reduced data staleness lag from nearly two days down to under a minute (typically 10-30 seconds).
  • 02Achieved a massive +1.67% NDCG offline ranking quality boost, a highly significant improvement for an established search system.
  • 03Demonstrated clear business impact with an approximate 0.33% increase in uncancelled bookings in live A/B tests.
  • 04Built as reusable, general-purpose platform features (Push Mode, NRT Model Transform) within open-source Chronon.
  • 05Eliminated the need to rewrite or duplicate complex transformer inference logic for streaming pipelines.

RECOMMENDATION

When scaling real-time recommendation systems, engineers should leverage standardized ML feature platform features like Chronon to enable near-real-time inference rather than building ad-hoc, use-case-specific streaming pipelines.

The Problem

Airbnb's previous search ranking model relied on daily batch jobs to update embeddings from guest journeys, leading to an data staleness lag of up to two days and failing to capture immediate search intent during active planning.

The Solution

Airbnb integrated two new general-purpose capabilities—'Push Mode' and 'Near-real-time (NRT) Model Transform'—into its open-source ML feature platform, Chronon, enabling instantaneous model inference and embedding updates triggered by streaming guest activity.

The Result

The update latency for guest embeddings was reduced from up to two days to under a minute (typically 10 to 30 seconds), yielding a +1.67% NDCG improvement in offline evaluations and a 0.33% increase in uncancelled bookings during online A/B testing.

Trade-off

Shifting from batch to near-real-time streaming inference introduces higher operational complexity and streaming compute costs, demanding robust monitoring to ensure reliable event streaming sequence and pipeline health.

03

Key Concepts

Concept · 01

Chronon

Airbnb's open-source machine learning feature platform designed to unify batch and streaming feature engineering and serving.

  • Acts as the backbone for managing feature definitions and updating low-latency feature stores.
  • Serves as the foundation where the new Push Mode and NRT Model Transform capabilities were added.
Concept · 02

Push Mode

A Chronon feature where streaming jobs publish a lightweight notification as soon as a feature value is committed, acting as an event-driven trigger.

  • Triggers immediate downstream model inference as soon as a guest's recent search or listing views are logged into the system.
Concept · 03

Near-real-time (NRT) Model Transform

A platform capability in Chronon that allows deployed ML models to be executed directly within the streaming pipeline data flow.

  • Runs the Transformer-based sequence encoder model directly within the streaming flow to produce and store fresh embeddings without waiting for daily batch jobs.