Jump to content

User:Accraze/MachineLearning/MiniKF

From Wikitech
MiniKF is no longer used for development. See https://phabricator.wikimedia.org/T293677 for more info.

MiniKF

MiniKF is a single-vm distribution of the Kubeflow suite of tools running on minikube. We are using this distribution for multiple sandbox clusters for Lift Wing development.

System requirements

  • 12GB RAM
  • 2 CPUs
  • 50GB disk space

Sandbox Details

KFv1.1 sandbox

This sandbox is used for initial service development on the v1alpha2 inference services.

  • 8 vCPU
  • 32 Memory(GIB)
  • Inference namespace: kubeflow-user
  • Auth: Dex
Software Version
k8s v1.16
Kubeflow v1.1
KFServing v0.3.0
Istio v1.5.7
Knative v0.11.2

KFv1.3 sandbox

This sandbox is used for multi-model deployments and benchmark load tests. We also use it to run v1beta1 inference services.

  • 16 vCPU
  • 64 Memory(GIB)
  • Inference namespace: kubeflow-user
  • Auth: Dex
Software Version
k8s v1.16.5
Kubeflow v1.3
KFServing v0.5.1
Istio v1.9.0
Knative v0.14.3

Monitoring

MiniKF comes preconfigured with Prometheus metrics collection and Grafana for building dashboards and viewing metrics. You can expose the Grafana app by doing a port-forward in a new terminal:

kubectl port-forward --namespace knative-monitoring \                                                                                                                                                    
        $(kubectl get pod --namespace knative-monitoring \                                                                                                                                                   
         --selector="app=grafana" --output jsonpath='{.items[0].metadata.name}') \                                                                  
         --address 0.0.0.0 8080:3000

Now you should be able to access the dashboards at <your-IP-address>:8080