Alibaba Cloud’s Agent-Ready Big Data AI Infrastructure: Boosting Data Development from Hours to Minutes

Why Enterprises Need Agent-Ready Data Infrastructure

Industry forecasts indicate that 85% of companies will internally deploy agents in the next two years, causing a massive increase in per‑interaction data volume that traditional big‑data and AI stacks cannot handle. Agents require high‑concurrency calls, cross‑engine analysis, and sub‑second interaction latency.

Core Architecture of an Agentic System

The system is organized into three layers: the Infrastructure Layer (compute resources, storage, data engines, and AI tools), the Agent Platform Layer (runtime management, orchestration, development framework, and external tool integration), and the Security Layer (Responsible AI governance). The key distinction from classic Retrieval‑Augmented Generation (RAG) is that agents possess both a "brain" (model) and "hands and feet" (tooling) to execute complete workflows.

Five Core Components

Intelligent Data Lake : Unifies unstructured and structured data representations, enables SDK‑based multi‑engine access, and supports elastic resource pools to replace fixed clusters.

Real‑Time Processing : Guarantees minute‑level (or sub‑minute) latency, provides staged feedback for multi‑round dialogs, and replaces hour‑level batch pipelines.

Vector Retrieval Upgrade : Expands from thousand‑dimensional text embeddings to ten‑thousand‑dimensional multimodal vectors (video, audio, text, tables), supports insert‑and‑query and mixed‑modality search, reducing token and compute costs.

Model Service Infrastructure : Offers zero‑downtime deployment, elastic scaling, PD (Prefill‑Decode) separation, KV‑Cache acceleration, and supports diverse GPU hardware (including domestic GPUs) for embedding, rerank, and foundation models.

Security Governance : Implements permission control, data‑quality validation, cost‑control via computation reuse, data‑lineage tracking, privacy compliance, and auditability for dynamic agent‑driven pipelines.

Use Cases

Case 1 – Internal Model‑Training Data Platform : Built on the Agent‑Ready stack, it stores both open‑source and proprietary datasets, provides real‑time retrieval, no‑code pipeline definition, and fine‑grained access control, enabling seamless data ingestion and model training.

Case 2 – Taobao Flash‑Sale Data Development : By adopting the ALEC solution plus DataWorks Data Agent, the traditional half‑day to two‑day workflow was compressed to 5‑10 minutes, with humans only handling final validation and release.

FAQ Highlights

What is Agent‑Ready Big Data AI Infrastructure? A data architecture designed for agents, consisting of a unified lake, real‑time processing, multimodal vector search, elastic model serving, and security governance.

Key challenges? Multi‑component technical complexity (Spark, StarRocks, Elasticsearch), high GPU hardware cost, and stringent data‑quality and compliance requirements.

Difference from RAG? RAG performs a single knowledge lookup per query, whereas agents involve tool calls, external environment interactions, and result verification, generating orders‑of‑magnitude more data and demanding second‑level response times.

Core products in Alibaba’s solution? DLF (managed lake supporting Paimon and Iceberg), Sub‑Agent‑exposed multi‑engine SDKs, DataWorks Data Agent for skill encapsulation, and integration with external frameworks such as OpenAI.

Real‑world impact? Data development time reduced from hours to minutes, with measurable efficiency gains in both internal model training and e‑commerce flash‑sale pipelines.

Overall, the Agent‑Ready architecture provides a foundational, secure, and high‑performance data backbone that prepares enterprises for the upcoming era of intelligent agents.