How Alibaba’s Tmall International Revamped Its Main‑Image “Sticker” System for Scale and Flexibility

The main product image (referred to as a “sticker”) is a critical conversion driver on Taobao and Tmall, presenting key product information such as price, discounts, and selling points across search results, recommendation lists, and detail pages.

Original workflow involved four parties: suppliers supplied base assets (transparent backgrounds, scene images, text points), product operators scheduled promotions in the inventory system, the UED team designed marketing‑aligned templates, and the sticker system automatically composited the final image for deployment.

2023 iteration was driven by rapid business growth, prompting a major system upgrade (nicknamed “sticker”). The new system had to support:

Automatic composition to improve operational efficiency.

Real‑time up/down of discount images to avoid stale promotions.

Fine‑grained, category‑specific templates to increase conversion.

The upgraded architecture introduced standardized point‑of‑sale data, integration with the marketing middle‑platform (UMP), and an expanded domain model comprising three core elements—selling points, templates, and placement slots—plus four abstract domains: product points, sticker images, sticker placement, and sticker scheduling.

Key architectural goals were business isolation and customisation while retaining shared core code. The solution combined a single application with multiple deployment groups, achieving isolation at several layers:

Infrastructure: separate Redis, MySQL, ADB instances per business.

Deployment: multi‑group deployment of the same code base.

Permissions: distinct UI pages, HTTP APIs, RPC providers, and role‑based checks.

Data: isolated storage for each sub‑business.

Scheduling: map‑reduce style multi‑instance processing.

External dependencies: shared core services (e.g., product lookup) with business‑specific extensions via second‑party packages.

High‑concurrency scenarios were addressed in two ways:

Point‑change bursts: MQ notifications trigger sticker generation; consumer throttling and retry mechanisms smooth traffic spikes.

Internal scheduling bursts: SchedulerX2 provides map‑reduce and API capabilities for task production, distribution, execution, and retry.

Migration strategy employed a dual‑write approach where both old and new systems processed operations, with the old system completing the final placement. Data reconciliation ensured consistency before fully switching over. A two‑step gray‑release plan—first data reconciliation, then operational migration—allowed rollback at the product level if anomalies arose.

Price‑consistency assurance was added via an OCR‑based price inspection service that scans generated stickers, flags mismatches, and triggers alerts for rapid remediation, thereby protecting consumer experience and reducing support costs.

Conclusion – Since early 2023, the sticker system has evolved into a platform that supports business isolation, customisation, multi‑image multi‑placement experiments, and automated price verification, delivering higher conversion rates and a more reliable shopping experience.