FMCG case study · MLOps

A Cost-Efficient MLOps Platform for Deep Learning on Azure Kubernetes Service

DS Stream built an MLOps platform for deep learning for a global FMCG / CPG company. The client had to scale deep learning model training and inferencing for high traffic and large image datasets. A web application on Azure Kubernetes Service with custom autoscaling, GPU memory sharing and GitHub Actions CI/CD led to faster development cycles and lower operating costs.

Talk to our MLOps team See all case studies

Key results

Lower operational costs and faster development cycles for the client's deep learning model operations.
Custom Kubernetes autoscaling was a more affordable and flexible option than Azure Machine Learning (AML) managed endpoints.
GPU costs went down because memory sharing among workers reduced underutilization.
High traffic and large image datasets are handled by a platform that scales cost-effectively on Azure Kubernetes Service (AKS).
CI/CD pipelines with GitHub Actions made testing, validation and deployment seamless, so the client can iterate quickly.

What problem did the global FMCG / CPG company need to solve?

The client needed to scale its deep learning models to handle high traffic and large image datasets efficiently. Data scientists also needed simple access to these models. They had to test and deploy models quickly without managing the underlying infrastructure themselves.

What the client asked for

Scalability for high traffic and large image datasets
Simple access to deep learning models for data scientists
Fast testing and deployment of models
No infrastructure complexity for the data science team

What did DS Stream build on Azure Kubernetes Service?

DS Stream deployed a web application on Azure Kubernetes Service (AKS) that gives data scientists easy access to deep learning models. The platform was designed on Azure infrastructure with three goals: scalability, cost-efficiency and ease of use for the data science team.

Key components

Automated CI/CD that follows the client's organizational standards and best practices
Horizontal and vertical resource autoscaling to make better use of workers and GPUs
Model versioning, monitoring and retraining pipelines so models adapt to data drift and concept drift
Trained ML models containerized as microservices to optimize inference for online and batch predictions

How did the platform reduce the cost of deep learning model training and inferencing?

Two design choices lowered costs. Custom Kubernetes autoscaling replaced Azure Machine Learning (AML) managed endpoints as a cheaper and more flexible option. GPU memory sharing among workers reduced underutilization, so the client paid for less idle GPU capacity.

Delivery and iteration

CI/CD pipelines built with GitHub Actions automate testing, validation and deployment. This lets the client release model updates quickly and deliver more value to its users. Read more about our approach to model deployment and serving.

Which technologies power the platform?

The stack combines open-source ML tooling with Azure infrastructure:

Languages and ML: Python, PyTorch
Containers and orchestration: Docker, Kubernetes, Azure Kubernetes Service (AKS)
Cloud: Azure
Delivery: CI/CD with GitHub Actions
Data and messaging: PostgreSQL, ServiceBus
Observability: OpenTelemetry

This setup supports scaling deep learning on Kubernetes while keeping the platform simple for data scientists to use.

FAQ about this MLOps project

Why use Kubernetes autoscaling instead of Azure Machine Learning managed endpoints?

In this project, customized Kubernetes autoscaling on AKS was a more affordable and flexible alternative to Azure Machine Learning (AML) managed endpoints. Horizontal and vertical autoscaling let the team tune worker and GPU utilization to the client's traffic and dataset sizes.

Does the platform support both online and batch predictions?

Yes. Trained ML models are containerized as microservices. This design optimizes model inference for online predictions and for batch predictions, so the same platform serves both use cases for the client's data scientists.

How does the platform keep models accurate over time?

The platform includes model versioning, monitoring and retraining pipelines. Together they help models adapt to data drift and concept drift, so data scientists can update models as the underlying data changes instead of relying on a fixed version.