How Volcano Engine’s Multimodal Data Lake Tackles AI Agent Challenges

By 2025, AI agents have become a leading focus in the industry, driving enterprises to enhance operational efficiency and decision intelligence. Volcano Engine’s Data Agent aims to meet these demands, but transforming a generic conversational tool into a precise, business‑aware intelligent agent requires high‑quality data, deep domain knowledge, structured context management, and real‑time user feedback loops.

These requirements raise the standards for underlying data infrastructure, especially in storage formats, compute engines, and data management paradigms.

The rapid growth of unstructured data in China—projected to increase from 51 ZB to 129 ZB within five years, with 80% being non‑structured—creates three core challenges for agent‑centric data infrastructure: (1) storage format limitations for multimodal data, (2) compute engines optimized for CPU rather than GPU‑heavy AI workloads, and (3) data management models that cannot handle the scale and diversity of unstructured assets.

To address these, Volcano Engine launched a next‑generation multimodal data lake solution built on five pillars: a multimodal storage layer using Iceberg and Lance formats (plus MCAP and LeRobot for embodied AI), a unified training‑inference data processing platform, upgraded Spark integrated with Ray, Daft, ByteHouse, and Flink, rich data processing operators (e.g., PDF extraction, image embedding, audio separation, video key‑frame extraction), and comprehensive management tools for versioning, exploration, and sharing across modalities.

The solution introduces the Processing Agent, enabling natural‑language invocation of underlying operators and compute frameworks, dramatically lowering the barrier for unstructured data handling.

Lance, a columnar table format designed for AI, supports object storage (TOS), HDFS, and vePFS, offering hot‑cold data tiering, efficient column addition, schema evolution, and transparent compression that can halve storage costs. Volcano Engine integrated Lance deeply with schema management and version control.

Daft, the chosen compute engine, provides native multimodal data types, Python DataFrame and SQL interfaces, a Rust‑based vectorized execution engine, heterogeneous GPU‑CPU scheduling, and lightweight distributed execution via Ray and Flotilla, enabling high‑performance processing of massive multimodal datasets.

Through these innovations, Volcano Engine’s multimodal data lake supports petabyte‑scale deployments, seamless GPU‑CPU collaboration, efficient data loading, large‑scale shuffle handling, and unified multimodal retrieval, positioning it as a comprehensive solution for AI‑driven enterprises.