Design

Rafay's Kubeflow based MLOps platform is an "enterprise-ready" MLOps platform based on Kubeflow tightly integrated with a number of related add-ons and community software. It is heavily optimized for specific infrastructure types/providers resulting in an extremely streamlined deployment and ongoing lifecycle management.

The architecture diagram represents the Rafay's MLOps Platform that is designed and optimized to be operated as a cloud native application inside a Kubernetes Cluster. In this example, the underlying Kubernetes cluster is based on Rafay's MKS Distribution.

---

Infrastructure

At the core of the architecture is a Kubernetes Cluster that is either provisioned/imported using Rafay in a Project in your Org. Data is persisted and accessed through MinIO, while relational data is managed via MySQL, and Redis is deployed for online feature storage. Within the Kubernetes cluster, Istio is used for managing the service mesh and traffic control. Feast, a feature store, is integrated into the platform for managing and serving machine learning features.

User Components

The platform is designed to be used by data scientists, researchers, and MLOps engineers. They are provided access to Kubeflow, a machine learning toolkit, along with its associated components such as Jupyter notebooks for interactive computing, Kubeflow Pipelines for orchestrating machine learning workflows, and MLflow for tracking experiments and model lifecycle management.

Authentication

The platform ensures secure access and user authentication. This is performed using Dex for managing authentication within the Kubernetes environment.

---

High Level Steps

The image below describes the simple steps that an administrator has to follow to configure and deploy the entire software stack into their Kubernetes cluster.

---