Building a One‑Stop Machine Learning Platform for Meituan Delivery

0. Introduction

AI has become a core focus for internet companies, and Meituan Delivery faces increasing order volume, rider count, and complex scenarios that demand faster, better, and more accurate machine‑learning algorithms. The article shares the team’s experience building a one‑stop ML platform to address these challenges.

1. Business Background

By July 2019 Meituan took over 30 million daily orders, operating the world’s largest on‑demand delivery network. The platform must handle ETA prediction, intelligent dispatch, map optimization, dynamic pricing, and more, while balancing experience, efficiency, and cost.

2. Evolution of the Meituan Delivery ML Platform

2.1 Why Build a One‑Stop ML Platform

Existing commercial and open‑source solutions either focus only on model training or require extensive engineering effort. A unified platform lets algorithm engineers concentrate on strategy rather than infrastructure.

2.2 MVP Stage

Initially each business line built its own “silo” tools, enabling rapid iteration but causing duplicated effort and inconsistent feature definitions.

Repeated wheel‑building : separate feature engineering, model training, and online prediction pipelines.

Inconsistent feature definitions : duplicated features with differing statistics hindered collaboration.

2.3 Platform Stage

To avoid duplication, an algorithm engineering group created a unified platform with the following components:

Resource scheduling on Hadoop/YARN, integrating Spark ML, XGBoost, and TensorFlow, with extensibility for other frameworks such as Meituan’s MLX.

Visual offline training UI that abstracts framework differences and generates DAG graphs.

Model management for registration, discovery, deployment, versioning, and high‑availability online prediction.

Offline and real‑time feature platforms for large‑scale feature computation.

Version management and A/B testing services.

3. Turing Platform Details

3.1 Offline Training Platform

Provides a drag‑and‑drop visual interface to build training pipelines, supporting custom parameters, auto‑tuning, and custom loss functions. It also generates an MLDL (Machine Learning Definition Language) file that bundles preprocessing logic with the model for consistent offline‑online execution.

3.2 Model Management Platform

Supports both local deployment (Java library embedded in business services) and remote deployment (dedicated online prediction cluster accessed via RPC), handling model registration, discovery, switching, and downgrade.

3.3 Offline Feature Platform

Transforms massive daily delivery logs into offline features stored in Hive, then serves them to online services. To meet high‑concurrency demands, features are grouped into KV pairs to reduce lookup overhead.

3.4 Real‑Time Feature Platform

Collects and processes streaming delivery data (rider location, load, road conditions, merchant status) to generate up‑to‑date features for immediate model inference.

3.5 A/B Experiment Platform

Implements a three‑stage workflow—pre‑experiment AA grouping, real‑time AB traffic splitting, and post‑experiment effect evaluation—tailored for the interdependent delivery ecosystem, including time‑slice based splitting to mitigate sample scarcity.

4. Summary and Outlook

The Turing platform now supports multiple Meituan business units, enabling algorithm engineers to focus on model strategy and significantly improving productivity. Future work includes deeper support for deep learning, more flexible Python model definitions, and a decoupled online prediction platform.