How Manbang Built a Cloud‑Native Real‑Time Data Platform with Flink & Hologres

Overview

Manbang Group’s logistics platform uses mobile internet, big data and AI to create a smart logistics ecosystem. The real‑time data platform is built on a cloud‑native architecture that combines Alibaba Cloud Flink and Hologres, providing a unified data bus for drivers, shippers and operations.

Platform Architecture

The system separates production (real‑time) and offline clusters across different data centers. Real‑time data is ingested via Flink CDC into Hologres, forming a real‑time supply‑demand layer that augments traditional ADM layers with driver distribution, cargo distribution, secondary cargo categories, and contextual information such as weather and epidemic status.

Real‑time features (driver behavior, cargo status, route pricing, etc.) are computed at minute‑ or second‑level and fed to algorithms and operational services.

Real‑Time Decision Workflow

The decision platform follows six steps: data insight, intelligent attribution, real‑time alerting, crowd profiling, OLAP‑based strategy development, and rule‑engine integration, culminating in A/B testing.

It supports page ranking, search ranking, intelligent recall, push distribution, driver membership and pricing strategies, all powered by real‑time features.

Migration to Cloud‑Native Flink

Manbang migrated 560 Flink jobs from a self‑built cluster to Alibaba Cloud’s fully managed Flink service in 1.5 months. SLA improved from 95 % to 99 %, operational staff reduced from three to one (saving 420 person‑days per year), and development cycles shortened by two days per task.

Resource consumption dropped 35 % (average 6.67 CU per job to a 40 % reduction), and the platform now benefits from cloud‑native elastic scaling and isolation.

Real‑Time Data Warehouse & Feature Engine

The warehouse integrates user, cargo, vehicle, traffic, transaction, finance and risk data, with a focus on driver‑cargo location state. Data is synchronized via Flink CDC and enriched through real‑time calculations, forming a “real‑time supply‑demand engine”.

Batch feature computation framework supports multiple time windows (seconds to hours) and complex derived features (percentiles, rankings, cross‑features such as weather and epidemic). This reduced feature development time from three days to two, and later to under one day, while consolidating 120 tasks into 16, saving 200 Core resources.

Products Built on Real‑Time Data

Two main products were launched: a real‑time metric alerting platform (“Beacon”) and a real‑time data service platform that exposes APIs via a unified “data supermarket”. Both enable proactive monitoring and easy data consumption for downstream services.

Future Plans

Roadmap includes Real‑Time Decision Platform 2.0, large‑scale Flink‑on‑AI, and a cross‑cloud OLAP engine built on Hologres and other cloud services.

/ END /