Lotte ON's Journey Building Its Own Real-Time Image Processing Service, LLIS

Lotte ON built LLIS (Lotte Live Image Service) in-house to reduce the high licensing costs and complex structure of external solutions. The core idea was to change the existing always-on server-based real-time conversion structure into an on-demand architecture centered on AWS S3 Object Lambda, and to cache the conversion results long-term in CloudFront.

LLIS is broadly divided into three servers.

• llis-pipeline: Handles product image registration and data synchronization, collecting changes by linking internal Kafka events with DynamoDB Streams and Kinesis.
• llis-image: Processes images in real time into the required specifications and responds when requested by users.
• llis-goodsdetail: Refines image quality and security in product spec HTML, automatically applying HTTPS conversion and Lazy Loading.

The request flow is simple. When a product is registered, images are stored in S3, and when a customer opens the product detail page, CloudFront calls S3 Object Lambda to instantly convert the image. The first generated response is stored in the cache, and subsequent identical requests are returned directly from the edge cache without going through Lambda again.

To improve cache efficiency, a Virtual Caching Key and long TTL were applied, and the system was configured to reuse identical results based on parameters such as width, format, and quality. As a result, they achieved a 30%+ reduction in S3 Object Lambda invocation count and an average image response speed reduction of about 100ms.

Image conversion was implemented by having Lambda read the original file and use Sharp to perform resizing, format conversion, and compression. In addition, Access Point-based control was used to allow only internal calls without an externally exposed endpoint, enhancing security.

In the product spec area, http resources were changed to https, and where HTTPS access was not possible, they were replaced with proxy URLs to eliminate Mixed Content issues. On the operations side, CloudWatch, S3 log storage, and Slack notifications were combined to automate incident response, achieving an average 50% reduction in incident response time and an operational automation rate of 80% or higher.

As a result, LLIS became an internalization case that simultaneously achieved cost reduction, operational simplification, and maintained real-time processing performance. Lotte ON presented a direction to further advance its Serverless and Edge-based structure going forward, reducing repetitive operations and concentrating development resources on more valuable areas.