How JD’s Advertising Architecture Scaled for 11.11: Lessons in Cost‑Cutting and Performance

Background

During the 11.11 shopping festival, the advertising division faced a massive surge in traffic, handling over 100 billion external playback requests—far beyond previous peaks.

Retrieval Service Scaling

Within two weeks, retrieval services across advertising systems were expanded with double‑shard capacity. Per‑shard memory usage dropped 30‑40%, and the 99th‑percentile latency (TP99) improved by 6‑25 ms, reducing memory pressure and cutting machine costs by roughly 30%.

Model System and Strategy Optimizations

The team optimized five key subsystems across four ad platforms (primary focus, search, recommendation, external). Improvements included shard‑level ad‑slot granularity, model architecture refinements, intelligent service protection, index data safety checks, feature‑log degradation, finer‑grained monitoring, and automatic traffic‑switching in anomalies. A dual‑model, second‑level real‑time switching strategy doubled the traffic capacity of the recommendation homepage, preventing revenue loss from dropped requests.

Data‑Processing Systems Enhancements

New billing infrastructure became lighter and ten times more hardware‑efficient. The rebuilt real‑time statistics system, leveraging Flink’s high‑throughput, low‑latency capabilities, saved five times the hardware resources and markedly improved stability. The upgraded order‑tracking 3.0 system tripled processing capacity, delivering timelier data. The refreshed MTA system enables advertisers to evaluate marketing effectiveness at a fine granularity.

Cloud Migration and Resource Management

Since April, the technology efficiency team migrated core workloads to the cloud, increasing allocated cores from 40 k during the 618 event to over 110 k, representing more than 50 % of the group’s total cloud cores. Over half of the self‑built Spark/JStorm/Flink clusters were also moved to the cloud, alleviating on‑premise compute shortages and paving the way for larger‑scale migrations.

Advertising Platform Integration

The customer‑platform team prepared the “Tian‑Gong” and “Sou‑Ke” systems to deliver high‑quality traffic with excellent user experience. The “Jing‑Zhun‑Tong” ad‑delivery system underwent database query optimization, data migration, and sharding, boosting throughput. The “Jing‑Sou‑Ke” programmatic interface integrated Baidu search resources, enhancing keyword‑based bidding. Creative assets were generated using an animation library for rapid 11.11‑specific templates. Reporting, backup, and pre‑audit services were also re‑engineered for high availability without additional machine resources.

Overall, the coordinated effort across architecture, algorithms, operations, and cloud migration ensured a stable, high‑performance advertising ecosystem capable of handling the unprecedented demand of the 11.11 promotion.