MLRun

MLRun is an open-source Machine Learning Operations (MLOps) orchestration framework (machine learning operations) for building, deploying, and managing data and Artificial Intelligence (AI) pipelines on Kubernetes-based and cloud-native infrastructure.

• End-to-end MLOps orchestration for data ingestion, training, deployment, and monitoring (MLOps framework).
• Serverless runtime and job execution on Kubernetes, including functions, workflows, and automation (container orchestration / serverless compute).
• Integrated feature store for managing and serving production-grade data features for models (feature store / data management).
• Support for real-time and batch pipelines, model serving, and online/offline feature access (data pipelines and model serving).
• Pluggable integrations with common data science tools, data sources, and Continuous Integration and Continuous Deployment (CI/CD) systems (ecosystem integration / DevOps).

More About MLRun

MLRun is an open-source MLOps orchestration framework (MLOps framework) created by Iguazio for automating and managing the lifecycle of data science and Machine Learning (ML) workloads on cloud-native infrastructure. It targets teams that run production workloads on Kubernetes and need a programmable layer for packaging code, data, and configuration into reproducible, automated pipelines.

The project provides an abstraction called functions (serverless compute) that encapsulate containerized workloads such as data preparation, training jobs, model serving endpoints, and batch processing tasks. These functions run on top of Kubernetes, allowing teams to define runtimes, resources, triggers, and parameters as code. MLRun supports workflows (workflow orchestration) to connect multiple functions into directed pipelines, enabling automated execution from data ingestion through training, testing, and deployment.

A central capability in MLRun is its integrated feature store (feature store / data management), which manages the lifecycle of ML features from ingestion and transformation to storage and serving. The feature store supports both offline and online feature storage, enabling reuse of features across projects and consistency between training and inference. This addresses common enterprise requirements around data versioning, governance, and operational reuse for ML features.

MLRun includes mechanisms for model deployment and serving (model serving), including real-time and batch serving options. It enables configuration of model endpoints, monitoring, and scaling using Kubernetes-native constructs. Through its function abstraction and configuration-as-code approach, MLRun integrates with CI/CD pipelines (DevOps / CI/CD), supporting automated builds, tests, and deployments of ML applications.

The framework interoperates with Python-based data science tools (data science tooling), common data sources, and storage systems, as documented by Iguazio. It exposes a programmatic Software Development Kit (SDK) and APIs (developer tooling / APIs) for defining projects, artifacts, runs, and pipelines, which allows integration into existing enterprise engineering workflows, notebooks, and automation scripts.

In enterprise and institutional environments, MLRun is used as a control plane for MLOps on Kubernetes, providing repeatable packaging of workloads, versioned tracking of runs and artifacts (experiment tracking / metadata management), and managed execution across clusters. It fits into categories such as MLOps platforms, feature store platforms, workflow orchestration tools, and Kubernetes-based serverless frameworks, and is positioned to support deployment on-premises (on-prem), in the cloud, or in hybrid setups as described by Iguazio.