Technical Overview of JD.com Search and Recommendation Systems for the 11.11 Shopping Festival

JD.com has built a self‑developed distributed search engine that provides precise, sub‑second results across multiple entry points such as mobile apps, PC, WeChat, and international sites. To handle growing traffic and data, a new architecture was rolled out in stages, aiming for full deployment by early next year and supporting the critical 11.11 sales peak.

The search team leverages deep‑learning techniques to improve ranking fluidity, mitigate the Matthew effect, and capture rapid sales spikes through timeliness models, which have already boosted conversion rates and GMV during the October rollout.

Similarly, JD's recommendation system integrates multi‑feature, multi‑dimensional user behavior and context to deliver intelligent product suggestions. Component‑based design enables rapid configuration across many platforms (APP, mini‑programs, PC, etc.), allowing the team to meet numerous 11.11 requirements efficiently.

For the promotion, dedicated deep‑learning models were created to enhance recall relevance and model real‑time user actions, further increasing conversion and GMV during the event.

In preparation for 11.11, the search and recommendation teams conducted extensive cluster stress tests, failure simulations, and downgrade plans. Automated operations platforms handled rapid scaling and secure deployments, while testing teams generated million‑level traffic for full‑stack validation, leading to hundreds of rehearsal cycles.

Beyond technical work, the department organized a "Programmer Festival" with team‑building activities, gifts, and morale‑boosting events to support the staff during the intense preparation period.