Baidu Cloud Native Data Platform: Empowering Enterprise AI in the LLM Era

This article is adapted from a presentation at Baidu Cloud Smart Summit 2024 - Cloud Native Forum, discussing how enterprises can build effective AI capabilities in the large language model (LLM) era.

Large models are trained primarily on publicly available internet data, lacking enterprise-specific internal data and industry knowledge. This limits their ability to handle real enterprise business problems. The industry has two main approaches to solve this: fine-tuning and Enterprise RAG (Retrieval Augmented Generation), enabling LLMs to leverage enterprise data combined with general models to achieve enterprise intelligence.

Building good enterprise intelligence faces several challenges: (1) Enterprise data exists in various storage systems including structured and unstructured data, requiring processing including collection, cleaning, transformation, and annotation before being usable by LLMs or vector databases; (2) As businesses grow, data scales simultaneously, requiring higher performance and cost-efficiency from data platforms; (3) LLM business requires high agility, making platform usability critical.

Baidu Cloud introduced significant updates to their database and big data products:

PegaDB : A self-developed KV database competing with open-source Redis, with features including batch loading, cross-region multi-active capability, and cold-hot data separation. It offers 30-50% lower pricing than competitors, supporting over 1 million QPS with p999 latency under 10ms.

GaiaDB 5.0 : The latest cloud-native relational database version with HTAP capabilities, supporting both columnar indexes and columnar engines. It introduces compute node scale-out for distributed cloud-native integration, and offers Serverless version achieving over 50% compute and 80% storage resource savings.

Vector Database 2.0 : Self-developed vector database for RAG scenarios, with 2.35x improved memory utilization and 7x performance improvement over open-source alternatives. Includes AI Search SDK for quick knowledge base application development.

Palo 2.0 : Real-time analysis database based on Apache Doris, with TPC-DS performance improved over 10x and storage cost reduced by over 80% through cold-hot separation.

Platform capabilities include:

DBSC 2.0 : One-stop database DevOps platform supporting development, management, and security auditing for over 10 database engines including MySQL, GaiaDB, Redis, and openGauss.

EDAP 2.0 : Lakehouse integration platform providing comprehensive data lake capabilities with full support for unstructured data governance, deep integration with AI platforms, lakehouse management supporting Iceberg, Hudi, and DeltaLake formats, and end-to-end Serverless computing.

DBStack : Database privatization platform supporting multi-cloud and hybrid cloud deployment, providing complete database capabilities including open-source, cloud-native, and commercial databases.