JD Retail’s Unified HDFS Storage: Cross‑Region and Hierarchical Storage Practices

Overview

With the rise of the big‑data era, JD Retail built a robust big‑data platform where Hadoop Distributed File System (HDFS) serves as the core distributed storage layer, providing high reliability, scalability, and the foundation for downstream analytics and upstream data ingestion.

Cross‑Region Storage Challenges

Rapid business growth exposed limitations of single‑datacenter deployments, leading to issues such as insufficient disaster‑recovery, inconsistent metadata across sites, excessive data redundancy, and uncontrolled inter‑datacenter links.

Cross‑Region Architecture

JD adopted a full‑copy plus full‑network topology: every DataNode reports to a common NameNode, enabling unified metadata management, consistent data placement, and seamless migration without service disruption. The new design boosted migration efficiency by 350 % and read‑write performance by over 70 % through read‑only nodes and write‑weight balancing.

Data Flow and Asynchronous Replication

Cross‑region data flow is driven by cross‑region tags; when a client writes to its local region, the NameNode issues a CR‑check task that asynchronously copies blocks to other regions, ensuring consistency and reducing latency.

Asynchronous Updater and Flow Control

An asynchronous updater processes backlog tasks with priority queues, preventing large tables from monopolizing resources. Flow‑control modules split block‑copy queues per region and apply dynamic rate‑limiters to protect inter‑datacenter links.

Hierarchical Storage Strategy

To address hot‑warm‑cold data management, JD classifies storage nodes into SSD (hot), HDD (warm), and high‑density HDD (cold) tiers. Data is automatically migrated based on access patterns monitored by an LRU‑based access monitor, with tier‑management modules creating conversion tasks that are scheduled by a distributed task manager.

Automatic Tier Conversion

Hot‑to‑warm and warm‑to‑hot migrations use internal data‑move mechanisms, while cold data is converted to erasure‑coded (EC) blocks via TTL‑driven lifecycle policies, improving storage efficiency and durability.

Performance and Cost Benefits

The combined cross‑region and hierarchical storage solution increased overall system throughput by roughly 10 %, raised EC coverage to 30 %, and reduced cold‑data storage costs by 90 %.

Practical Use Cases

Two scenarios illustrate the impact: (1) cross‑region lifecycle management automatically moves stale data to EC‑based cold storage, cutting redundancy; (2) data‑scheduling leverages tiered placement and task‑drift to balance compute resources across regions, enhancing job latency.

These innovations demonstrate how JD Retail optimizes large‑scale distributed storage for both performance and cost efficiency.