XuanYuan An AI-Native Database
1 minute read ∼ Filed in : A paper noteQuestions
This is almost impossible to implement and to tune.
Introduction
Background & Motivation
Stand-alone database: data storage, data management and query processing; PostgreSQL, MySQL
Cluster Database: high availability and reliability: DB2 and SQL server.
Distribured Databases (cloud-native database): elastic computing and dynamic data migration
Gap
Exisitng DB cannot support various applications and diversified computer power.
- AI for DB
- DB for AI: design in-database machine learning frameworks, which utilize DB techinques to accelerate AI algorithm
- GPU hardware integration
Goal
The paper try to propose a DB system design and challenges in providing follow properities.
- Self-configuring, self-optimizing, self-monitoring, and self-diagnosis etc.
- Provide AI capabilities using declarative languages
- Utilize diversified computing pwoer to support data analysis and ML.
DB4AI
AI as UDF
Model can be embeded in the DB and we could provide UDF or stored procedures for each algorithm. Then use can call UDFS or SPs to use AI algorithm.
AI as Views
Make the trained AI algorithm as a view, which is shared by multiple users. The model can be then updated offline.
Model-free AI
Database can automatically recommend the algorithms fir the user scenarios.