Advances in Alibaba's Advertising Engine: Serverless Architecture, Recall, Strategy, and Creative Technologies

Alibaba Mama, the benchmark for e‑commerce advertising in China, has progressed through several rapid innovation phases: the 2019 "three‑horse carriage" era (Direct Train, Diamond Booth, Super Recommendation), the 2020 explosion of Super Live, Super Interaction, and Fast Push, and the 2021 shift to a full‑domain auto‑driving era represented by Wanxiang Tai.

The advertising engine, as the core system that handles traffic, implements product functions, and satisfies client demands, now faces new challenges: diminishing traffic dividends, the need for green, low‑carbon compute, and deeper digitalization of merchants. Consequently, the engine is undergoing a fundamental evolution.

From a traditional service‑oriented iteration model to a Serverless paradigm, the engine confronts three main pain points: scattered R&D focus due to the need for developers to understand physical deployment details; coarse delivery granularity that hampers rapid feature rollout; and high difficulty in cross‑module reuse because each team builds highly customized frameworks and data schemas.

The proposed solution consists of three unified abstractions: a unified runtime that hosts PaaS capabilities and enables DataFlow‑based execution with operator fusion and parallelism; a unified data abstraction that treats all engine data as standardized Tables, making data scheduling transparent to developers; and a unified business abstraction that standardizes operator inputs/outputs and reduces operator granularity, allowing operators to “drift” across modules and stages, thereby improving latency, data handling, and compute efficiency.

The recall engine has evolved from keyword‑based methods to vector‑based and, most recently, model‑based recall. Its architecture is abstracted into three logical roles—MatchServer (orchestrates the overall recall flow), Trigger (generates recall signals), and Searcher (retrieves candidate ads). To overcome the limitations of online two‑stage recall, an offline‑precomputed, large‑scale, non‑truncated recall system (TFMS) has been built, enabling seamless sharing of code and resources between online and offline pipelines.

The strategy engine has progressed from single‑stage bidding to multi‑stage, multi‑objective automatic bidding, incorporating advanced auction mechanisms such as Deep GSP and Neural Auction. The end‑to‑end strategy platform comprises three subsystems: an offline training system (supporting reinforcement learning and hyper‑parameter optimization), a real‑time control system (providing low‑latency data warehousing and model inference), and an online bidding system (executing various bidding and auction algorithms).

The creative engine leverages AI to generate high‑quality image, text, and video ads at massive scale. It consists of a creative production engine that processes raw assets into millions of candidate creatives daily, and a creative delivery engine that selects and assembles the best creative for each traffic scenario, ensuring native appearance and optimal performance.

Overall, Alibaba Mama’s advertising engine showcases a Serverless‑first, cloud‑native architecture, a flexible recall framework, an end‑to‑end strategy platform, and an AI‑driven creative system, all aimed at improving iteration efficiency, reducing latency, and delivering greater business value.