Kurly Migrates Airflow Operations to Kubernetes… Achieves 50% Cost Reduction and Stability

Kurly's Data Service Development Team successfully migrated Apache Airflow for its data pipelines from a managed service to a Kubernetes (K8S) environment and achieved stable operation. This migration was carried out with the goals of improving technical maturity, strengthening infrastructure capability through direct problem-solving, and reducing costs.

Key Operational Issues and Resolution Process

During the migration, CPU overload on the Scheduler and OOM (Out of Memory) issues on the Worker occurred when using GitSync. The CPU overload from GitSync arose because the new and development Airflow instances were syncing the same Git repository, and this was resolved by adjusting min_file_process_interval and analyzing the serialization load. CPU usage increased from 0.4 core to 2.7 core, but was stabilized through appropriate parameter tuning.

The OOM issue was addressed by analyzing K8S's 3-stage mechanism (OOM, Kubelet, OS). Container-level Request/Limit settings and node isolation using Taint & Tolerations prevented resource contention and improved incident response speed. Additionally, log loss caused by BigQuery delays and hitting the Concurrency ceiling was resolved by scaling out the number of Workers from 1 to 2, eliminating the bottleneck.

Results and Implications

Compared to the fixed costs of the managed service, using K8S nodes reduced operating costs by approximately 50%. Through the transition to a self-managed environment, the team strengthened its technical capabilities in Airflow and K8S, while also carrying out stability-securing work such as advancing monitoring. This confirmed that hands-on experience gained through trial and error is a key driver of infrastructure growth.