Design and Optimization of JD Advertising Retrieval Platform: Adaptive Compute Allocation, High‑Efficiency Search Engine, and Platform‑Scale Infrastructure

JD's advertising retrieval platform serves billions of users by performing the initial matching of users, goods, and ad spaces, facing the core challenge of balancing limited compute power with massive data volumes.

System Overview – The platform converts advertiser demands into a language understood by the playback system and performs large‑scale recall, returning hundreds of items from billions of candidates while considering user experience, advertiser goals, relevance, and platform revenue.

Main Line 1: Beyond Serverless – Data‑Driven Adaptive Compute Optimization – A distributed execution graph driven by RPC calls replaces rigid service dependencies, enabling global compute‑optimal scheduling and achieving over 16% latency reduction. Compute allocation, optimization, and iteration efficiency are addressed through a unified configuration center and one‑stop experiment platform.

Main Line 2: Racing Against Time – High‑Efficiency Retrieval Engine – The platform evolves from rule‑based recall to dual‑tower models with ANN, then to business‑aware hierarchical indexing, full‑library PQ indexing, and deep indexing based on EM, continuously improving recall speed and relevance while supporting arbitrary target optimization (e.g., eCPM).

Main Line 3: Platform Power – Platform‑Scale Infrastructure – By modularizing business logic into atomic operators (OPs) with clear data dependencies, the system achieves high observability, configurability, and plug‑and‑play strategy customization, supported by a massive experiment capacity and one‑stop configuration management.

Elastic System – A PID‑based elastic controller aligns CPU usage with target levels, while later stages allocate idle compute to maximize revenue, using value‑estimation functions trained on uplift metrics.

Future Outlook – Ongoing work will continue to enhance compute allocation, retrieval efficiency, and iteration speed across the three main lines, further scaling JD's advertising capabilities.