How AliExpress Boosted Orders by 5% Through Full‑Stack Performance Optimization

Theoretical Foundations

The article begins by introducing the AliExpress e‑commerce system and uses bounce‑rate calculations to motivate a full‑stack optimization approach. Conversion rate is defined as 1 – bounce_rate. As latency increases, bounce‑rate rises and conversion drops, illustrating performance loss. By modeling traffic distribution across latency intervals, the theoretical maximum traffic for each page path can be estimated, enabling precise measurement of the impact of performance improvements.

Platform Design

The platform is built independent of specific business domains, providing visualized performance metrics, real‑time monitoring, and full‑link tracking. All business units can plug into the platform, which aggregates data continuously to produce accurate measurements that guide whether optimizations should be rolled out globally.

Performance Optimization Process

Multiple business modules collect user behavior data (request times, connection times, etc.). Collected data are processed by a pipeline offering offline and real‑time analysis, with results stored in a business database and cached for quick retrieval. The backend supports scenarios such as alerts, monitoring dashboards, and reporting.

Implementation of Optimizations

Dynamic Acceleration Before optimization, dynamic requests travel from the user through the ISP to the origin server, incurring long latency. After optimization, dynamic content is pushed to edge nodes, reducing round‑trip time. Requests can be parallelized and CDN‑chosen paths further minimize latency. Challenges include preserving real user IPs and protecting the origin from attacks.

Static + ESI Static assets are cached at edge nodes, while dynamic content is fetched from the origin only when necessary. This reduces cache misses and speeds up content delivery.

Element Request Merging Multiple sub‑elements on a page are combined into a single request to the server, which then distributes and aggregates the responses, cutting down TCP handshake overhead.

CDN Scheduling Optimization By analyzing latency from different geographic nodes, the system routes users to the most efficient edge node. Tests showed Brazilian users experienced lower latency when served from Brazilian edge nodes compared to US or Argentine nodes, informing smarter CDN scheduling.

Business Results

The combined theoretical analysis, platform deployment, and optimizations yielded a 5.07% increase in orders, noticeable reduction in performance loss, and enhanced purchasing power across the system.

Architectural Reflections

The design encourages innovation that can be quantified and compared across contributors.

Technical democratization allows anyone with ideas to experiment and contribute.

Accurate measurement of individual performance contributions amplifies team impact.

Building trust step‑by‑step is essential for widespread adoption.