Olive Young Reveals 'Intentional Over-Engineering' Case Study: Implementing CDC and Kafka in a Low-Traffic SMS System

Olive Young revealed a case study where they intentionally introduced CDC and Kafka while refactoring an SMS/LMS sending system handling less than one message per second. The existing legacy daemon suffered from structural debt, including performance degradation due to resource sharing with the WAS, lack of standard monitoring, and reliance on verbal rules for preventing duplicate sends. However, since changing legacy code across dozens of services was impossible, a transitional structure was required.

Strangler Fig Pattern and Event Interception

The system adopted a structure applying Martin Fowler's Strangler Fig pattern, detecting INSERTs on the target table via CDC and passing them to Kafka without modifying the legacy code. This serves as an Event Interception mechanism that intercepts a table shared by dozens of services, a strategy designed to reduce the risks of gradual migration and realize early value.

Worker Design Considering Irreversible Sends

To address the irreversibility of sending and the asymmetry of DB transactions, the approach was applied where Kafka messages contain only IDs, and a separate worker retrieves the original data from the DB to send. Judging that 'sent but not recorded' has a higher recovery possibility than 'recorded but not sent,' the history is saved after sending, and duplicate sends are prevented through exclusive partition assignment in the Consumer Group.

Performance Isolation and Observability

Although the average sending volume is less than one per second, spikes exceeding 10,000 messages occur during specific times. The Kafka queue buffers these sudden spikes, and the isolated application blocks the impact on WAS performance. Measurements from the last 15 days showed an average processing time of 67ms per worker, securing enough capacity to handle more than 10 times the usual sending volume.

Over-Engineering as an 'Investment'

The transitional architecture explicitly defined its validity period by setting the removal point as '0 INSERTs into the target table.' Abstraction relative to expected changes is categorized as an 'investment,' while complexity that remains in a form that fails to deliver value is categorized as 'debt.' Olive Young concluded that accepting temporary complexity for the safety of migration is cheaper than the risks of a big-bang migration.