Kakao Automates DevOps with AI-Based Deployment Management Platform 'KakaoRelease'

Kakao has built and is operating KakaoRelease, an AI-based deployment management platform, to overcome the limitations of complex service dependencies and manual reviews. This platform enhances stability by collaborating with AI throughout the entire process, from deployment planning and execution to monitoring.

Optimizing Automated Deployment Note Generation

Previously, GitHub and Jira data were input directly into LLMs, causing context overload and quality variance issues. To address this, a three-stage pipeline was applied: extracting core information from data to summarize via Summary AI, and generating drafts in parallel via Generate AI.

• Data Compression: Jira and GitHub information was compressed by 95%, from approximately 44,000 tokens to 2,300 tokens, reducing costs and latency.
• Parallel Processing: By splitting notes into units of 4 for parallel processing, the average generation time was reduced by 80%, from over 120 seconds to under 30 seconds.
• Consistency Assurance: Few-shot Learning and RAG-based schema application maintained consistency in output format and tone.

AI-Based Risk Review and Monitoring

KakaoRelease stores approximately 500 past incident cases in a vector DB and analyzes risks by searching for similar risk patterns to the current deployment plan in real-time. It uses a hybrid approach where algorithms precisely verify policy violations and schedule conflicts, while AI provides flexible reasoning for alternatives.

For 24 hours post-deployment, Change Tracking functionality linked with Matrix AI detects changes in error rates, response times, and traffic patterns in real-time. This reduces the burden of manual monitoring for developers and aims for an autonomous operation system that analyzes causes and determines rollback necessity when anomalies occur. In the future, Kakao plans to complete E2E automation by integrating with various deployment tools like ArgoCD and Jenkins via a plugin approach.