Inside JD.com’s Self‑Developed Switch: Zero‑Touch Deployment for Massive 11.11 Sales

JD.com’s 11.11 global shopping festival achieved a record 2715 billion yuan in orders, highlighting the need for stable, high‑capacity networking. To meet this demand, JD Cloud introduced a self‑developed white‑box switch in 2020, which has been deployed online and operated reliably during the event.

Traditional data‑center switches are closed‑box devices from multiple vendors, lacking unified standards and requiring separate integration for hardware, software, configuration, monitoring, and management.

JD’s switch decouples hardware and software, custom‑orders hardware, selects ASIC chips, and builds its own operating system, JDNOS, based on the open‑source SONiC project. It offers a unified interface, automated deployment, and fine‑grained monitoring.

The network lifecycle—from construction to decommission—benefits from automation: when a device powers on, it automatically retrieves and installs the OS version, loads configuration, performs self‑checks, and reports to the management platform. The platform can perform automatic network cut‑over and supports hot‑reboot (warm‑reboot) upgrades without traffic interruption.

In addition to traditional CLI configuration, the switch provides a gRPC interface with a YANG model, enabling batch configuration changes and greatly improving efficiency over manual SSH commands.

Zero‑Touch Provisioning (ZTP) automates device onboarding, eliminating manual OS installation and configuration loading, thereby reducing labor costs and accelerating deployment.

The warm‑reboot feature allows process‑level, container‑level, or full‑system upgrades while keeping the forwarding plane active, ensuring uninterrupted service during version upgrades.

When using BGP, JD enables the GR feature to keep routes stable during control‑plane restarts. To address slow convergence in certain failure scenarios, a BGP + BFD solution provides sub‑second route recovery.

Traditional monitoring (SNMP, CLI, syslog) uses pull methods. JDNOS adopts telemetry push, allowing the platform to subscribe to data such as interface traffic statistics, CPU/memory usage, power, fan status, and optical module metrics, delivering real‑time, high‑efficiency data collection.

JDNOS also implements millisecond‑level traffic monitoring, pushing per‑second statistics to the platform, enabling precise visibility into traffic spikes during large‑scale events.

Network anomaly monitoring covers packet loss, queue congestion, forwarding latency, and path tracing, with alerts for hardware issues like bus errors, register faults, DoS attacks, and table errors.

Real‑time monitoring of optical modules tracks voltage, current, optical power, and temperature, providing predictive failure warnings and link‑stability alerts.

Regular consistency checks verify interface status, VLANs, ARP tables, routing, and ACLs; detected issues trigger alerts and automatic process repair without disrupting service.

By offering full‑stack monitoring beyond traditional black‑box capabilities, JD’s self‑developed switch has achieved zero recorded failures and continues to evolve with new features to meet growing bandwidth and functional complexity demands.