[if kakao 2022] Designing an ML Model Training Pipeline Using an MLOps Platform

Kakao Pay built an MLOps platform to solve the duplicate development problem and operational complexity of having to build separate serving servers and feature tables for each model. The platform consists of three core components: feature store, model training pipeline, and model serving.

The model training pipeline is divided into model training Workflow and model registry. AWS Step Functions was adopted as the workflow platform. Airflow lacked ML optimization, and AWS SageMaker Pipelines placed a burden on operators to deeply understand model code, but Step Functions supports diverse workflows and makes templating easy via the Python SDK.

For the model registry, AWS SageMaker Model Registry, which has lower operational difficulty, was chosen. The decisive factor was that it reduces management resources compared to the open-source MLFlow. Through this, an environment was established where data scientists can focus solely on model code, while MLOps operators can manage pipelines with standardized templates.