Introducing Coupang's Machine Learning Platform

Coupang's ML Platform is designed to speed up ML development used across search, pricing, logistics, recommendations, and advertising. The goal is to standardize the entire ML lifecycle—exploration, data preparation, model development, and production deployment—so developers can focus on the model itself instead of repetitive work.

There are three core problems. First, boilerplate for writing and training models needs to be reduced to shorten time to production. Second, CI/CD needs to be incorporated into ML development to reduce technical debt. Third, resources like GPUs, large-scale storage, and networking for distributed training need to be operated efficiently across a hybrid setup that mixes on-premises and AWS.

The platform's built-in services are managed notebooks, pipeline SDK, feature-store, model training, and model inference. Developers launch Jupyter notebooks using standard or custom containers to experiment, and use a Python SDK to query data, integrate with the feature-store, and build training and inference pipelines. Standard Docker images include libraries like TensorFlow, PyTorch, Sklearn, Hugging Face, Transformers, reducing dependency complexity.

Feature engineering is handled by a feature-store based on Feast. The offline feature store is used for sharing features for training and for model training, while the online feature store fetches features with low latency during inference and also serves as a prediction response cache for models with high computational cost.

The model training stack is framework-agnostic. Pipelines written by users are containerized and run on a Kubernetes cluster, with a batch scheduler placing them on appropriate hardware. Users can choose CPU and GPU types, and distributed training strategies like distributed data parallel and fully sharded data parallel are also supported to speed up training for large models.

On the operations side, Seldon is used as the inference platform on top of Kubernetes. Seldon integrates with TFServing, Triton, and custom Python wrappers, covering various model frameworks, runtimes, and CPU/GPU serving. Each model is deployed as an independent service, handling real-time traffic reliably through autoscaling, validation testing, canary deployments, and monitoring.

Among the success stories introduced are training Ko-BERT to improve search query understanding, and performing real-time price forecasting for products. The platform team benchmarks frequently used model architectures internally and shares trainer tuning and best practices to boost reusability and efficiency across the whole team.