Kubeflow Controller

Kubeflow controller is the open source re-implementation of part of the internal tool, ml-executor, in CaiCloud. This project is temporary and we hope we could use tensorflow/k8s powered by the Kubeflow community eventually to replace this controller.

Background

Recently Google open sources a new project Kubeflow, which is dedicated to making using ML stacks on Kubernetes easy, fast and extensible.

We have a similar project in CaiCloud and it serves for several years. In that project we took different approaches but similar design decisions to Kubeflow. Then we re-implement our internal tool and open source it to accelerate the development of the community powered controller tensorflow/k8s.

Overview

We separate Kubeflow into two parts:

Kubeflow controller, which contains the controller for Kuberflow CRD.
Kubeflow clientset, which contains the clientset and CRD specification for Kubeflow.

There are some differences between tensorflow/k8s and this controller:

caicloud/kubeflow-controller follows the design pattern of controller, which does not keep a state machine in the code and just moves from the current state to the desired state. tensorflow/k8s follows the design pattern of operator.
caicloud/kubeflow-controller uses pods to run the workers and PS, while tensorflow/k8s uses jobs.

Besides these, there are also some differences between tensorflow/k8s CRD and caicloud/kubeflow-clientset CRD:

We do not support TensorBoard in the CRD, since the TensorBoard and Serving support will be in other CRDs.

These three points are the biggest differences, others could be ignored or unified.

Get Started

The controller is easy to use, see docs/get_started.md for more details.

CONTRIBUTING

Feel free to hack on the controller! docs/development.md will help you to get involved into the development.

Acknowledgments