Kakao Recommendation Team Operates Large-Scale Platform Processing 1.7 Billion Recommendations Daily Across 15 Services

The Kakao Recommendation Team provides personalization and related recommendation technologies to 15 services, including KakaoTalk, Daum, Kakao Gift, Melon, and Piccoma. As of July 2022, the team processes an average of 1.7 billion requests per day through approximately 200 recommendation APIs, maintaining an average response time of 30–50ms and a maximum of 100ms by building a proprietary pipeline.

The recommendation algorithms utilize Contextual Bandit, Collaborative Filtering, Sequential Recommendation, CTR Prediction, and Text/Image embedding. The platform was developed based on Python, with some algorithms implemented in C++ for real-time computation optimization. The architecture follows an Event-driven architecture and uses various data storage solutions such as MongoDB, HBase, RocksDB, SSDB, and HDFS alongside Apache Kafka.

Instead of short-term click-through rates, the Recommendation Team is advancing the system to target long-term service growth metrics such as revisit rates. Additionally, the team is developing a SaaS-based platform to allow non-experts to easily utilize recommendation technologies, contributing to the expansion of the technical ecosystem by releasing open-source projects like N2 and Buffalo and operating the Kakao Arena competition.