Inside JD.com's 11.11 Tech: Cloud, AI, and Ops Strategies

Zhang Chen: 1500+ Emergency Plans, 150 Drills for 60‑Second Recovery

The team prepared over 1,500 emergency response plans, each application having a primary plan to ensure any failure can be resolved within 60 seconds. About 150 stress‑test drills, including joint exercises across lines, simulated real‑world traffic spikes, targeting the ability to handle 20 times the normal load.

Despite previous successes in 618 and past 11.11 events, Zhang emphasized that confidence must be backed by data‑driven rehearsals. JD.com operates one of the world’s largest Docker clusters (≈60,000 nodes), enabling rapid scaling for flash‑sale traffic and reducing operational overhead. Multi‑city data centers provide disaster‑recovery capabilities, while the JD Brain project drives personalized user experiences.

Operations and monitoring are critical; a real‑time dashboard displays the health of all systems, and mobile monitoring complements it. Data quality is paramount—JD’s extensive, accurate e‑commerce data fuels AI models that can increase GMV by billions of yuan through precise marketing and inventory decisions.

Liu Haifeng: Full Cloudization, 100k Docker Scale and Operations

Liu oversees JD’s public and private cloud architecture. The private cloud, launched early this year, consolidates storage, middleware, and elastic compute resources. For 11.11, elastic compute will handle 100% of traffic, enhancing disaster recovery, operations, and rapid fault handling.

A new data center supports up to 100,000 simultaneous Docker instances (up from 11,000 at 618). Docker‑based deployment cuts approval and packaging steps, dramatically shortening release cycles. Elastic compute can expand thousands of containers within 10 seconds, ideal for flash‑sale spikes.

Network infrastructure uses 10 GbE hardware with VLAN+OVS on OpenStack. Liu notes the need for skilled staff to build and maintain such infrastructure and stresses that cloudization is an ongoing, multi‑phase effort to support future growth.

Wang Xiaozhong: Multi‑Transaction Centers and Hot‑Standby

The “multi‑center transaction” project distributes traffic across several regional centers, each acting as a primary or secondary node. Real‑time data buses synchronize master data (products, merchants, users) to sub‑centers and aggregate transaction data back to the master.

The first phase, using the Langfang data center, provides hot‑standby capacity for 20× the previous peak traffic. Hot‑standby improves throughput and availability, while cold‑standby handles overflow when needed. Challenges include ensuring data consistency and scaling large clusters.

Wang illustrated the end‑to‑end flow from product listing to order placement, emphasizing that both hot and cold backup strategies are essential for resilience and cost optimization.

Yang Guangxin: JD Brain’s One‑Two‑Three‑Four

JD Brain, launched earlier this year, focuses on AI‑driven personalization across PC and mobile platforms, covering about 60% of orders during 618. Its four layers are data, model, system, and application.

High‑quality e‑commerce data enables detailed user, product, and community profiling. Hundreds of features feed machine‑learning models that predict ranking, demand, and seasonal trends, supporting millions of predictions per second.

The system delivers high‑performance inference at scale, while applications use these insights for targeted recommendations, precise marketing, and intelligent inventory management. Ongoing work includes improving cold‑start handling and exploring deep‑learning‑derived features.