Skip to content

Latest commit

 

History

History

20240321-kubecon-paris

README.md

KCP ML Shop demo

This is a demo of a simple machine-learning service, hosted on KCP. It shows how to use the KCP as a platform to deploy a machine-learning model and expose it as a REST API. Use global shards to run models close to data (geo placement).

Demo recording: https://youtu.be/7op_r9R0fCo?si=8KxUnEgSASBRJCFG&t=1515

Pre-requisites

  • KCP running with shards - shards root and beta are required.
  • Kubectl installed
  • Kubectl-workspace plugin installed

Demo scenario

Platform operator persona

The platform operator is responsible for setting up the platform and enabling 3rd party. This is done by creating APIExports and APIBindings in dedicated workspace. Workspace management is done via kubectl-workspace plugin.

export KUBECONFIG=kcp.kubeconfig
  1. Show current "platform backend" view:
k ws tree
  1. Bootstrap the assets. This shows the platforms provider enabling 3rd party to handle the management of certain resources via APIBindings:
go run ./cmd/controller/ init --kubeconfig=kcp.kubeconfig
k ws tree
  1. Show what server side looks like:
k ws use root:ml:training
k get apiexports
k get apiexport -o yaml

Now that a platform owner has pre-created assets, and 3rd party can use them to deploy models, let us move to service provider persona.

Important: In the demo we used the same kcp.kubeconfig for both platform operator and service provider. In real-world scenario, those would be different kubeconfigs, preferably generated by the platform operator and certificate-based.

Service provider persona

  1. The service provider has a controller to run pointing to this APIExport:
export KUBECONFIG=idp.kubeconfig

Important: Mount code is not yet open-sourced and it's a work in progress. The cluster running the controller has to have access to both APIExport exported URLs, showed in the previous step. In a real-world scenario, this would be done via VPN or other secure means, where shards can be accessed.

  1. Deploy the controller:
k apply -f ../kcp-ml-shop/demo/deploy/namespace.yaml
k create secret generic kubeconfig --from-file=kcp.kubeconfig -n ml
k apply -f ../kcp-ml-shop/demo/deploy/deploy.yaml

User persona

  1. User has a model to deploy - Create an ML-enabled workspace (in 2 locations):
k ws create ml-europe --location-selector=name=root --type ml-training
k ws create ml-us --location-selector=name=beta --type ml-training
  1. Deploy the model
cat demo/model.yaml
k ws use ml-europe
k create -f demo/model.yaml
# get to US:
k ws use ..
k ws use ml-us
k create -f demo/model.yaml