How Alibaba’s Search Engine Evolved Over a Decade of Double‑11: From Offline Models to Real‑Time AI

Alibaba’s technical team launched the "Ten Years of Code" series after the tenth Double‑11 event, inviting core engineers to review the evolution of search intelligence on the e‑commerce platform.

Four Evolution Stages

Stage 1: Emerging Power – Pora Streaming Engine

In 2014, analysis of Double‑11 data revealed that hot‑selling SKUs often ran out of stock while users still received traffic, leading to low conversion. The self‑developed streaming engine Pora collected all click, add‑to‑cart, and purchase logs, aggregated them by product, and joined real‑time inventory data to compute real‑time sell‑through and conversion rates, feeding the results back to search and recommendation engines. This enabled large‑scale real‑time computation for the first time, boosting PC and mobile revenue.

Stage 2: Real‑Time Dual‑Link System

From 2014 onward, Alibaba built an online learning pipeline (online + decision) that moved from batch offline models to online learning, allowing continuous model updates without caching all data. Techniques such as Multi‑Armed Bandit (MAB) and zero‑order optimization were applied to select optimal ranking strategies in real time, outperforming static offline‑trained models during the high‑traffic Double‑11 period.

Stage 3: Early Exploration – Deep + Reinforcement Learning

In 2015, online learning proved effective, but challenges remained: models over‑relied on cumulative signals from the zero point, and MAB’s discrete strategy space limited rapid adaptation. Alibaba introduced a parameter‑server‑based online learning framework, building pointwise conversion‑rate estimators and pairwise matrix‑factorization models, and deployed them via swift for simultaneous feature and model prediction.

Stage 4: Full Deep‑Learning Era

By 2017, the streaming platform was rebuilt on Blink/Flink , supporting 24‑hour uninterrupted processing and scaling machine‑learning jobs from dozens to hundreds. Powerful CPU/GPU heterogeneous services enabled large‑scale online deep learning, covering semantic search, multi‑modal product representations, online deep‑learning mechanisms, global ranking that accounts for item context, and multi‑scene collaborative decision‑making. These advances were recognized in KDD 2018, IJCAI 2018, and WWW 2018.

Key Drivers of Evolution

Dynamic e‑commerce environment requiring instant capture of product, price, and inventory changes.

Personalization since 2013, moving from pure query matching to query + user context + region + time.

Shift from PC to mobile, demanding real‑time modeling of fragmented user behavior.

Future Outlook

Despite impressive progress, current models still rely heavily on product tags and behavior logs, lacking deeper semantic understanding of user intent. Alibaba aims to combine human knowledge with machine intelligence to achieve cognitive AI, enabling search to anticipate diverse user needs and provide truly intelligent, “human‑like” experiences.