Kurly's Delivery Team Eliminates 'Fake Errors' to Cut Alarm Fatigue and Establish a Log Management System

Kurly's Dawn Delivery Product Team overhauled its log management system to address the alarm fatigue problem caused by increasing complexity in its last-mile service system. Previously, many 'fake errors' were logged indiscriminately without clear rules, causing vigilance to dull when real problems occurred.

Re-establishing Log Level Standards

Through meetings and data analysis, the team established clear standards for the ERROR and WARN levels. ERROR was defined as a situation requiring immediate response, configured to trigger alarms instantly, while WARN was set as a state that could become problematic if it occurs frequently, adjusted to trigger alarms only when a threshold is exceeded within a specific period.

• ERROR: Requires immediate response, includes unintended exceptions
• WARN: Frequency-based monitoring, common range applied (e.g., 100 times within 5 minutes)

Operational Strategy and Effects

For logs from external system integration timeouts or areas beyond the application's control such as Kafka and Redis, the team analyzed patterns and either set up alarms or excluded them. As a result, system error logs over the past 3 days were reduced to 4 cases and warning logs are managed at 355 cases, enabling near-real-time monitoring that detects problems before users do.

Logs must be managed continuously like code, and rather than achieving a perfect initial setup, it is important to flexibly adjust alarm frequency according to traffic and user patterns during operation. Establishing this culture has improved developers' debugging capabilities and strengthened service stability.