Introducing Kubeflow: Composable, Portable, Scalable ML Stack for Kubernetes | Hacker News Reader

Introducing Kubeflow: Composable, Portable, Scalable ML Stack for Kubernetes

blog.kubernetes.io

14 pointsTheIronYuppie11 comments

Anyone interested in test driving KubeFlow on Ubuntu, can follow this tutorial:

https://tutorials.ubuntu.com/tutorial/get-started-kubeflow

I am red hat and I can tell you that several of us are very excited about Kubeflow. We look forward to a productive 2018.

See our blog post on why we are doing this and how we hope to contribute: https://blog.openshift.com/machine-learning-openshift-kubern...

Thank you to all the folks who helped us launch this product and community, including (in no particular order) CaiCloud, Jupyter Labs, Red Hat & OpenShift, Heptio, WeaveWorks, Canonical, Container Solutions, CoreOS, Katacoda and many more!

Please let us know if you have any questions.

Disclosure: I work at Google on Kubeflow

Hey all, happy to answer your questions about Kubeflow on Ubuntu and our Distribution of Kubernetes (CDK).

(I work for Canonical)

Thank you !! More the merrier and happy to be here. Looking fwd.

Hi this is Alexis from weaveworks, just to say you can see how we support kubeflow here - https://www.weave.works/blog/kubeflow-and-weave-cloud

Congrats to the Kubeflow team. This is a really exciting new project for everyone trying to do ML on Kubeflow. I'm excited to see what this will mean for the Kubernetes in general given the close contact between the Kubeflow team and the Kubernetes team. I'd love to see some more data centric primitives coming out in new K8s releases.

Interesting point! Anything in particular? Specifically, we love the extensibility of the Kubernetes CRD and that has given us a lot of what we were looking for, and Hardware Acceleration as a schedulable resource is due to land very soon. But do tell!

Disclosure: I work at Google on Kubernetes

I have a lot of ideas about this. I'm one of the creators of Pachyderm, which does data pipelining on Kubernetes. The first concrete thing that I think K8s could have would be some sort of a notion of a DataSet. Right now the closest thing I think is a gitRepo volume. GitRepo volumes are different from most other k8s volumes in that they're not meant for persisting state, they're meant as a way to inject data (which might actually be code) into a k8s environment. I'd like to see that idea expanded, I could easily imagine a similar thing that allowed you to mount data from Object stores like s3 and GCS, that would do a lot to make K8s immediately useful for data workflows. Being able to output to volumes like this and pass them between various pods would be the next logical step.

This is an amazing idea! Have you thought about exploring to see if we can overload CSI to accomplish something like this? https://github.com/kubernetes/community/blob/master/contribu...

Hey all,

I'm the engineer at Google leading our Kubeflow effort. Happy to talk to anyone interested in think about contributing or using Kubeflow.

Reply on news.ycombinator.com