Arrikto: ML Model Deployments on Kubernetes Can Get Better

Data scientists and engineers typically work in parallel to DevOps production pipelines when they create machine learning (ML) models. Through so-called MLOps, the ML models these teams build are often then integrated into the application-deployment environment once ready for production. These data scientists or MLOps teams might also rely on GitOps to serve as the immutable application structure with a Git repository so that the ML models are distributed across different environments, such as hybrid clouds.

To facilitate the deployment of ML models and applications on Kubernetes, open source Kubeflow has emerged as one of the more promising open source tools to help with the process. Still, the challenges associated with MLOps teams deploying ML models and applications on Kubernetes — even with tools such as Kubeflow — are fraught with difficulty and complexity. This is due to both the data science-centric nature of ML-model development as well as the well-documented complexities associated with managing Kubernetes environments.

ML tool and platform provider Arrikto’s latest release of Arrikto Enterprise Kubeflow (EKF), featuring Kubeflow 1.4, offers improvements for data scientist teams working with DevOps and operations teams to integrate ML models on Kubernetes. It serves to further the software provider’s mission of supporting data scientists’ goals of building and deploying ML models faster, more efficiently, and securely, the company says.

Solve Machine Learning Challenges

“To solve [ML model development] challenges today, DevOps has to patch together a multitude of different tools, create deployment automation infrastructure and then re-do most of the work if they change environments — and still the workflow for the data scientists is not seamless,” Constantinos Venetsanopoulos, CEO and founder at Arrikto told The New Stack.”This is where Kubeflow and EKF come together to solve these challenges with a single platform and single user experience for the entire organization.”

This release enables data science and operations teams to collaborate more efficiently by “radically” reducing the time it takes to bring ML models to production,” Venetsanopoulos told The New Stack. Specific features Venetsanopoulos communicated include:

• Automation: For data scientists, a streamlined end-to-end workflow from experimentation to model with Kale, a workflow tool for Kubeflow that orchestrates all Kubeflow components.
• Reproducibility: For data scientists, SecOps, legal and DevOps, the versioning of notebooks and pipeline features include code, libraries, data and metadata using Rok, a data-management solution for Kubeflow.
• Security: For SecOps and DevOps, integrated secrets management to secure access to external services, integrating authorization providers such as Google, Okta, PingID, etc, while providing advanced isolation for users and teams.
• Portability: For the entire organization, GitOps-based automation tooling to deploy the same platform on Amazon Web Services (AWS), Azure and Google Google cloud provider environments and on on-prem infrastructure.

The company also communicated additional details on how EKF supports the left-to-right ML-model development cycle through GitOps. These GitOps features include how EKF was created to support an automated “end-to-end” deployment process, with a wizard that displays details about the infrastructure prior to deployment. EKF also offers a deployment automation tool that generates Kustomize manifests to help customize application configurations that meet the infrastructure requirements.

As mentioned above, EKF is also used to commit ML-model manifests to a Git repo to ensure immutability and reproducibility. The manifests are also used to deploy updates and configuration changes on Kubernetes.