How JD Cloud’s Self‑Developed Cloud Disk Powers Ultra‑Stable Mega‑Sale Events

JD Intelligent Cloud (JD Zhilian Cloud) has repeatedly supported JD’s massive sales events such as 618 and 11.11 by providing ultra‑stable cloud disk services. During the 618 promotion, nearly 30,000 cloud disks were used, storing about 5 PB of data, and the platform handled peak traffic smoothly.

The cloud disk is a core block‑storage resource offering low latency, durability, and high availability. JD chose a fully self‑developed approach, focusing on functionality, performance, and stability. After years of development, external customers now use the service more than internal ones.

It serves a wide range of workloads—including logistics, advertising, AI, online education, gaming, big data, and live streaming—providing extreme stability, high availability, and rich data‑management features, with a claimed 99.99999% data reliability.

The system is built on the Raft protocol with custom enhancements. Data is stored using a three‑replica strong‑consistency model. The control plane is centralized for better cluster control and resource utilization, while the data plane is distributed. The architecture consists of three main components: client, management node, and storage node.

Management nodes handle capacity management, status monitoring, and configuration changes. Storage nodes persist the three‑replica data, with devices deployed across different racks to isolate fault domains. Clients use a private protocol and can run on containers, virtual machines, or bare metal.

Fine‑grained resource management and defensive code design protect against hardware failures, network issues, and operational errors. User volumes are split into small replica groups distributed across racks. Each group has a Leader and two Followers; the Leader handles I/O and synchronizes Followers. Automatic failover occurs without affecting the business.

NVMe disks are used as a cache layer: client requests are served from NVMe for immediate response, while background processes migrate data based on hot‑cold classification, improving performance while preserving reliability.

The backend performs silent data scanning to verify consistency of cold data and can migrate it to cheaper storage, reducing costs.

Snapshots are implemented incrementally and stored in object storage, shared across availability zones. They support cross‑zone replication, delayed loading, and pre‑warming to accelerate snapshot‑based VM provisioning.

Smart NIC technology offloads storage networking, enabling zero‑copy RDMA transfers from the client to the storage backend, reducing CPU usage and latency—ideal for high‑performance databases and I/O‑intensive applications.

Looking ahead, JD Intelligent Cloud plans to continue advancing storage media, networking, and software stacks to deliver higher performance, lower cost, and greater stability.