Kubernetes as a hybrid solution for deploying complicated workloads, has made it possible for organizations to move complex workloads to the platform, while leveraging on rich APIs, reliability and performance provided by the container orchestration system.
And perhaps, the fastest growing use case is using Kubernetes as the deployment platform of choice for machine learning (ML), with the open source Kubeflow project to help in the deployment of machine learning stack on the Kubernetes container orchestration system.
While the building of production-ready ML system involves various components, and more often mixing vendors and different solutions; the management of these varied services for organizational setups brings additional barriers and complexity in the adoption machine learning.
Kubeflow toolkit, now in beta is intended to help in the deployment of machine learning workloads across multiple nodes, whereby breaking and distributing a workload adds up to computational overhead and complexity.
As Kubernetes is already tasked with making it easier to manage distributed workloads, Kubeflow is aimed at making the running of the workloads more portable, scalable, and simple.
Kubeflow includes scripts and configuration files, which users can customize with their configuration and scripts to deploy containers to any chosen environment.
The Kubeflow project open source Github repo is dedicated to making using ML stacks on Kubernetes easy, fast and extensible. And the repository includes: JupyterHub to create & manage interactive Jupyter notebooks and Tensorflow Custom Resource (CRD) to use CPUs or GPUs, as it's easily adjustable to the size of a cluster.
Kubeflow requires Kubernetes 1.8 or later, in a cluster and works with the Ksonnet framework, version 0.11.0 or later, for writing and deploying Kubernetes configurations to the clusters.
No comments