Ctrip Model Engine Platform: An Integrated End‑to‑End Service for Real‑Time AI Model Deployment

Author : Li Mei, Application Development Engineer in Ctrip’s Hotel Data Intelligence team, joined Ctrip in 2016 and has worked on cross‑recommendation and personalized hotel ranking services.

1. Introduction

Recent advances in artificial intelligence have been widely adopted in the internet industry, driving personalized and fine‑grained services. However, existing model deployment workflows still suffer from three main drawbacks: (1) manual data preparation, (2) reliance on offline prediction, and (3) long development cycles for real‑time model services.

To address these issues, Ctrip’s Data Services team launched the Model Engine Platform , a unified platform that reduces repetitive work, accelerates model rollout, and strengthens monitoring and evaluation of online models.

2. Platform Construction

The platform is built around four core modules—Feature Management, Engineering Management, Model Management, and Product Management—forming a closed‑loop from data processing to model deployment.

Design Goals

Serve product managers, data scientists, development engineers, and testing engineers through a panoramic service view.

Provide second‑level latency for real‑time data and millisecond‑level prediction responses.

Support a wide range of Ctrip business scenarios, including A/B testing, gray‑release, and multi‑model fusion.

Overall Architecture

The logical architecture consists of four layers:

Feature Management : Handles both offline and online feature preparation, offering a shared feature store, automated export to real‑time stores (Redis, ES, Aerospike), and unified monitoring of feature health (missing rate, freshness, activity).

Engineering Management : Supplies reusable components such as a Feature Bus for automatic local cache construction and a Code‑Gen service that generates model service skeletons for various model formats (PMML, Python package, XGBoost). It also captures input data for debugging and supports automatic model retraining.

Model Management : Manages model file upload, feature association, publishing, version rollback, and gray‑release validation. Automatic sanity checks and load testing ensure only verified models are deployed.

Product Management : Allows product managers to create and manage “scenes” (model prediction business units), link them to engineers, and visualize the end‑to‑end workflow, including support for sub‑scenes, parallel/serial/fusion/switch routing, and multi‑model fusion.

3. Main Components

Model Service : Generated by the Code‑Gen service, it encapsulates model invocation (PMML, Python, XGBoost), supports single‑ or batch‑level predictions, and includes built‑in monitoring and version management. It abstracts model upgrades and experiment version routing.

Feature Bus : A configurable XML‑driven component that builds local caches (heap + off‑heap) and synchronizes data from batch sources (HDFS) or streaming sources (Kafka, Qmq). It guarantees sub‑second update latency for real‑time features while shielding developers from storage‑specific code.

4. Summary & Outlook

The Model Engine Platform consolidates Ctrip’s experience in model deployment, offering end‑to‑end support for feature quality monitoring, model serving, and operational oversight. Since its launch in late 2017, it has enabled over ten projects (hotel ranking, image recognition, advertising, etc.) and now enters a second phase to expand the feature library, support more model types, and enhance online model quality assessment across additional business units.