Clean-Up Project: Reducing Resource Waste in KakaoPay's Kubernetes

KakaoPay faced difficulties in its Kubernetes environment where CPU and Memory resources were over-allocated beyond actual usage, increasing server costs, and where resource shortages occurred when migrating new services.

The existing VPA (Vertical Pod Autoscaler) had a constraint in that it could not be used together with the already-in-use HPA (Horizontal Pod Autoscaler). To solve this, KakaoPay pursued the 'Clean-Up Project' to develop its own resource recommendation service.

The project architecture consists largely of two modules.

• Resource Recommender: Fetches metrics from Thanos, calculates resource recommendation values, and stores them in MySQL.
• AgitBot: Announces the calculated recommendation values on Agit, the company's internal bulletin board, to encourage development teams to improve resource efficiency.

For calculating resource recommendation values, the method from Google's Autopilot paper was adopted. It produces sophisticated recommendation values using the histogram-based concepts of Lower/Target/Upper bound, a safety margin (margin-fraction), and a confidence correction coefficient (confidence_multiplier) based on the data collection period.

Test results confirmed a substantial reduction in server resource usage, delivering tangible cost savings.