A General Feature Production Framework for Meituan Delivery Ranking System

The article introduces a generic feature production framework designed to improve the iteration efficiency of Meituan's food delivery ranking system, which is driven by machine‑learning models (GBDT). Feature engineering is identified as the bottleneck for rapid model updates.

Feature statistics are the foundation: offline pipelines compute various statistical features (e.g., merchant sales, user category preferences) from raw exposure, click, and order tables stored in Hive. The framework abstracts the statistical pattern into three dimensions—statistical object, statistical dimension, and metric—allowing flexible configuration and time‑decay weighting.

Four main steps constitute the feature production pipeline:

Feature statistics (ETL or Spark‑based aggregation).

Feature push (mapping Hive rows to Domain objects, serializing them, and storing into a KV store).

Feature fetch (online service retrieves serialized data from KV and deserializes).

Feature load (online feature operators derive high‑level features from the raw ones).

The Spark‑based statistical engine supports custom dimension and metric operators, aggregation functions (sum, average, concat, ratio, quantile), and time‑decay weighting. Configuration files (Toml) define objects, dimensions, metrics, and operators, enabling rapid addition of new features with minimal code.

For feature synchronization, an ORM layer maps Hive rows to Java Domain objects, which are then serialized (JSON or Protostuff) and stored in KV with a prefixed key. A unified KvService handles serialization, deserialization, and KV read/write.

Online feature loading uses FeatureOperator annotations to declare required offline features ( @Fetchers) and produced online features ( @Features). A DataPortal caches fetched Domain objects, and a mapping Model → Feature → FeatureOperator → DataFetcher ensures efficient data reuse across multiple models.

In summary, by abstracting each stage of feature production, the proposed framework allows developers to add new features by defining a Domain class and a few configuration entries, dramatically reducing development effort and accelerating model iteration cycles.