Source | Xinyan Technology

文 | Jia Ningyu

On December 20, VLDB2024, the top international database conference, announced a new batch of papers. Alibaba Cloud's new technology PilotScope was successfully shortlisted. This platform technology can realize "one-click deployment" of AI algorithms in the database, greatly reducing the number of AI algorithms in the database. The application threshold has opened up a new path for database intelligence. On December 20th, the international top conference for databases VLDB2024 announced a new batch of papers, and Alibaba Cloud's new technology PilotScope successfully made it to the list. The platform's technology can achieve "one-click deployment" of AI algorithms in databases, greatly reducing the application threshold of AI algorithms in databases and opening up a brand new path for database intelligence.

Alibaba Cloud announced that it will open source all PilotScope technologies for free on the same day

Why is it difficult to make database intelligent?

Database is a basic software technology that is crucial to the national economy and people's lives. The continuous updating of database technology has an important impact on all walks of life in the digital era. One of the frontier areas is database intelligence (AI4DB, i.e. database intelligence)

The current database system is very complex and has very high stability requirements. Even just matching and debugging an AI algorithm with a database requires engineers from both parties to work closely for weeks or even months, which is inefficient and results in poor results

The more common situation is that AI engineers don’t understand the details of databases, database developers don’t understand AI, and the two fields don’t even know the programming language (AI development mostly uses Python, databases mostly use C/Java), it’s very difficult Easy to cause rupture.

Generally speaking, companies in the industry usually choose to embed some AI algorithms directly into the database to replace certain functional modules of the database, such as intelligent query optimization modules. However, this customized approach results in very high development, maintenance, and upgrade costs. Every time the AI ​​algorithm is upgraded and replaced, the development process needs to be redone. At the same time, changing the code base of the database will also bring additional risks

Because of this, despite the rapid development of artificial intelligence, the practical application of related results has not yet become popular in the database field

Is there a common platform technology that can more effectively apply artificial intelligence algorithms to databases?

This became the starting point for the thinking of the Alibaba PilotScope project team

PilotScope project leader Zhu Rong said: "AI4DB, AI and DB are both done by people, but the bridge at this connection, But it has never been done well. We want to build a public bridge between AI algorithms and databases to make communication between the two parties smoother."

Cross-tech innovation from 0 to 1

Zhu Rong described PilotScope as the "super administrator" of database AI. Through the PilotScope platform, AI engineers only need to focus on designing general AI algorithms to implement the deployment and application of different databases; while database users can call APIs like Likewise, the idea of ​​using AI conveniently and efficiently took about 2 years from conception to implementation. Zhu Rong said: "It involves the intersection of algorithms and systems, the intersection of AI and databases, the intersection of research and development, and the intersection of academia and industry. It is a true intersection of technology."

According to his introduction, the project After many rounds of polishing, the team finally developed a brand new middleware system platform. By abstracting and generalizing module and interface definitions at the database and AI system levels, the AI ​​algorithm can be implemented in the database within hours or even minutes. Key deployment", this is the current PilotScope

PilotScope is useful for parameter tuning, index recommendation, cardinality estimation, and query It provides more than 10 AI algorithms for mainstream database tasks such as optimization, and has successfully adapted to two mainstream open source databases such as PostgreSQL and Spark

According to experimental data, using PilotScope to embed AI algorithms into the database is faster than traditional The "hard implant" method can speed up tasks such as query optimization by 1 to 2 times. In addition, the additional cost of deployment caused by PilotScope itself is basically negligible, and the performance is excellent

PilotScope performs "micro-intrusion" on the database and introduces Intelligent detection, rollback, isolation and other mechanisms to reduce the risk of AI hallucinations and achieve intelligent improvement while ensuring database stability

Zhu Rong said that in the past, artificial intelligence engineers and database developers needed to continuously collaborate and refine, and it might take weeks or even months to ensure stability. "With the help of our PilotScope, it only takes a few hours or even dozens of minutes to go online for testing directly. This zero-to-one technological innovation greatly improves development efficiency."

Open source drives the industrialization of AI4DB Process

PilotScope paper results have been included in VLDB. The VLDB review believes that PilotScope's pioneering system design based on application scenarios will open up a new direction of database intelligence.

According to our understanding, VLDB is one of the three top international database conferences, and only includes reports on academia and industry every year. Practice new results that have important impact. It is an authoritative indicator of database technology. The 50th VLDB Conference is planned to be held in Guangzhou, China in August 2024.

Picture Note: Top Database Conference VLDB2024

Zhu Rong said, PilotScope related technologies have been freely open sourced on GitHub and Modelscope communities. The team hopes to incorporate more AI algorithms and a wider range of databases into PilotScope through the power of the open source community, and explore more AI4DB innovations with developers

At the same time, PilotScope has begun to deploy on Alibaba Cloud Conduct pilot applications internally to conduct corresponding tests for industrial deployment

Zhu Rong said that AI4DB can only generate value in a real production environment. We hope that PilotScope can truly realize this and help people from all walks of life. Improve the efficiency and effect of database intelligence

Please attach the open source address:

https://github.com/alibaba/pilotscope

The above is the detailed content of Databases usher in the AI ​​fast lane, Alibaba Cloud releases new open source technology PilotScope. For more information, please follow other related articles on the PHP Chinese website!