tryolabs / k8s-demo

k8s-demo

Example of a couple of use cases for kubernetes that could be useful in ML.

Demos

1. First demo shows how to train a pytorch model for MNIST as a Kubernetes Job. See train for more details
2. Second demo shows how to use canary deployments on a simple API. See api for more details

Local setup

1. Install docker
2. Clone this repo
3. Optionally but recommended:
   1. Install pyenv and run pyenv install
   2. Setup a virtualenv with python -m venv .venv and activate with source .venv/bin/activate
   3. Install dev requirements with pip install -r requirements.txt
4. Install kubectl

Cluster setup

I only tested this on minikube, for the first demo I used my local GPU on minikube which requires sudo for alternatives look into train.

Minikube setup

1. Install minikube
2. Use this script to setup a GPU-enabled cluster or a vanilla cluster.