Kyung Hee University Festival Platform

Why we offered to build it

The Student Council outsourced the festival website each year, but student engagement was low and the site did little to solve problems at the event. Our team proposed building and operating the service ourselves.

We wanted students to find performances and booths easily, pub operators to focus on their guests, and future teams to inherit both the system and its usage records.

My responsibilities

As development lead, I drove product and technical decisions from planning through implementation, deployment, and operations. My responsibilities included backend and infrastructure architecture, AWS and Terraform, delivery workflows, code review, observability, and incident response.

Load testing and bottleneck analysis

A read API’s p95 latency was about seven seconds in load tests. Rather than immediately increasing the connection pool, I compared execution with and without virtual threads under the same resource limits and pool size.

Under the same conditions, throughput rose from 16.6 to 45.5 req/s, and p95 latency fell from 6.98 to 2.40 seconds.

Increasing the pool later reduced throughput. Based on the measurements, I adopted virtual threads and kept the pool at four connections. Higher concurrency also increased external errors on the AI path, showing that throughput and admission control were separate concerns.

Controlled local Docker comparison: 0.5 vCPU, 1 GB, 40 virtual users, 150 seconds. These figures are separate from live festival traffic.

Read the experiment and results ↗

Three days in the field

Over three days, the platform recorded 1,634,031 requests and 17,308 estimated unique visitors. Cumulative pub table occupancy was 1,386 hours and 45 minutes.

The platform operated September 28–30, 2026. The 13 pubs recorded 648 table sessions, 167 waitlist registrations, and 156 admission calls. Festival information received 44,360 views.

Requests include pages, images, and API calls in the same deployment. Table occupancy sums the duration of individual table sessions across the 13 pubs.

Visits, information views, and admission calls exceeded our targets, but sticker verifications reached only 99 against a target of 300. Shipping a feature and generating participation were different problems.

Source: Adelante festival results report, October 2, 2026. Unique visitors are estimated from visitor identifiers. These are team-wide service outcomes, not individual results.