Jonghyeok Park

Notification delivery was scattered across API servers, admin tools, and batch jobs, so nothing enforced the external API’s rate limits consistently, and scheduled sends and bulk-failure tracking had to be rebuilt in every path. I owned the architecture design, gathering the scattered delivery paths into a single Redis/BullMQ queue with built-in rate limiting, delays, and retries, controlled by horizontally scalable workers—handling 30 million Kakao notifications in a single month.

Delivery history and observability data were separated from the delivery path, with archival and analytics pipelines on Firehose, SQS/DLQ, Lambda, and S3, defined and deployed with Pulumi.

A decade-old ASP/VB system kept its business logic locked inside 699 stored procedures, making it hard to keep pace with a growing service. I analyzed every procedure and rebuilt orders, payments, accounts, catalog, points, and SCM as a Spring Boot multi-module platform, completing the migration in two months. I built a server-side Conversion API that improved ad conversion efficiency by 168%, and the service grew 700% in revenue within a year, reaching 150K cumulative members and 30K MAU.

In-house fulfillment kept the next-day dawn-delivery order cutoff at 5 PM because of the previous carrier’s loading constraint. Moving to Kurly 3PL, I integrated order registration and cancellation, inventory lookup, and delivery eligibility checks, extending the cutoff to 7 PM.

A single screen needed 40–60 per-date delivery-schedule lookups that took 500 ms–1 s against the external API; the data was predictable by date and tolerated some staleness, so caching was the right fit. Redis alone still cost 50–100 ms in network round-trips and serialization, so I placed Caffeine L1 in front with Redis Pub/Sub invalidation for cross-instance consistency, serving L1 hits in about 10 ms.

A circuit breaker and fallback isolated external API failures so the ordering flow could continue without interruption.

Flink, JDBC, Spark, and other interpreters shared resources in one container, so one interpreter’s heavy workload could shake the performance and stability of everything else. I contributed to moving interpreters into separate Kubernetes Pods, dividing their resource and failure boundaries and enabling independent execution and scaling.

This work earned the Excellence Award (NIPA President’s Award) at the Open Source Contribution Academy. During the team’s contribution period, Zeppelin recorded a 10/10 community health score—the highest among Apache TLPs including Kafka and Spark—while user discussion and PR review activity grew 246% and 858%, respectively.

ECC’s formatter hook runs after every file edit, and each run took about 3.3 seconds—breaking the editing flow every time. I added project-aware Biome and Prettier detection, then shortened the execution path by preferring local binaries, removing duplicate checks, and invoking hooks in-process. In a local benchmark, runtime fell from about 3.3 seconds to 63 ms—52× faster.