Building a DBA Review Bot Based on Amazon Bedrock That Applies In-House DB Management Rules with AI

The Yogiyo DBA team built a DBA review bot that assists judgment based on in-house DB management rules, in order to reduce the review burden of frequent DDL changes in a microservices environment. When a development team registers a DDL and its application schedule in DBportal, it is automatically deployed after DBA approval, and this review stage was the key point determining stability.

Previously, DBAs had to manually check the rules for every request, and there was a high risk of omission in complex requests mixing multiple DDLs or cases involving CDC or Replication target tables. The review bot performs a first-pass analysis of such requests and summarizes the following:

• Whether the rules are complied with and risk points
• Whether additional verification is needed
• Identification of CDC / Replication target tables
• DDL examples and recommended directions that conform to the rules

As a result of operation, by pre-filtering the risky tasks that DBAs must definitely review, the burden of repeated checks decreased, and by presenting the reasons for rule violations along with recommended directions, the communication cost between DBAs and developers also decreased. Provided in the form of a Slack chatbot, it allowed rule inquiries to be handled immediately without having to separately search the Wiki, and the overall processing speed from DDL request to review completion improved by about 20%.

The implementation was built centered on Amazon Bedrock's Knowledge Base and Agent. To utilize internal rule documents as knowledge, documents were collected daily at 6 AM using the Confluence API and a cron job and stored in S3, after which the OpenSearch Vector Store was updated via a Bedrock ingestion job. Documents were split by DDL task unit, and correct and incorrect examples were included together to improve search quality.

The agent was designed not as a simple answer generator but as a coordinator assisting DBA review. Instructions were minimized, strong constraints were applied to keep the KB search results as-is, and the response template was also fixed. Queries were processed by separating them as follows:

• Whether it is a review request
• Whether an actual DDL is included
• Whether it is a CDC / Replication verification request

Currently, the review bot operates as a document-based first-pass review tool, and it cannot yet judge information requiring DB access, such as table size or index configuration. Going forward, expansion toward MCP-based metadata lookup and a multi-agent structure is being considered to further improve review accuracy and consistency.