https://github.com/katanaml/katana-skipper

Simple and flexible ML workflow engine
https://github.com/katanaml/katana-skipper

docker docker-compose ingress k8s katana kubernetes machine-learning orchestration pipeline tensorflow

Last synced: about 23 hours ago
JSON representation

Simple and flexible ML workflow engine

• Host: GitHub
• URL: https://github.com/katanaml/katana-skipper
• Owner: katanaml
• License: apache-2.0
• Created: 2020-09-19T14:39:31.000Z (about 4 years ago)
• Default Branch: master
• Last Pushed: 2023-10-04T07:27:33.000Z (about 1 year ago)
• Last Synced: 2024-10-29T03:19:23.713Z (16 days ago)
• Topics: docker, docker-compose, ingress, k8s, katana, kubernetes, machine-learning, orchestration, pipeline, tensorflow
• Language: Python
• Homepage: https://katanaml.io/
• Size: 6.75 MB
• Stars: 410
• Watchers: 11
• Forks: 80
• Open Issues: 0
• Metadata Files:
  • Readme: README-GKE.md
  • Funding: .github/FUNDING.yml
  • License: LICENSE

Awesome Lists containing this project

README

        
# Katana ML Skipper GKE

Deployment Guide for Google Kubernetes Engine (GKE).

## Author

Katana ML, Andrej Baranovskij

## Instructions

1. Push Skipper images to Docker registry, this registry should be accessible from GKE

2. Open GKE Cloud Shell, follow GKE instructions in Kubernetes setup wizard

![OCI](https://github.com/katanaml/katana-skipper/blob/master/gke-shell.png)

3. Install [NGINX Ingress Controller](https://kubernetes.github.io/ingress-nginx/deploy/#gce-gke) for GKE

4. Clone Skipper repo

```

git clone https://github.com/katanaml/katana-skipper

```

5. Edit rabbitmq/rabbit-statefulset.yaml file, change storageClassName to 'standard-rwo'

```

vim rabbitmq/rabbit-statefulset.yaml

```

6. Edit api/api-ingress.yaml file, remove 'host' element to configure Ingress with GKE public IP

```

vim api/api-ingress.yaml

```

7. There is no need to create Persistent Volume on GKE, it will be provisioned automatically by Volume Claim. Remove this line from kubectl-setup.sh:

```

kubectl apply -f services/trainingservice/trainingservice-pv.yaml

```

8. Edit services/trainingservice/trainingservice-pvc.yaml, change it to support dynamic provisioning for Persistent Volume, remove storageClassName

```

vim services/trainingservice/trainingservice-pvc.yaml

```

```

apiVersion: v1

kind: PersistentVolumeClaim

metadata:

  name: training-service-claim

spec:

  accessModes:

    - ReadWriteOnce

  resources:

    requests:

      storage: 500Mi

```

9. There is no need to create Persistent Volume on GKE, it will be provisioned automatically by Volume Claim. Remove this line from kubectl-setup.sh:

```

kubectl apply -f services/servingservice/servingservice-pv.yaml

```

10. Edit services/servingservice/servingservice-pvc.yaml, change it to support dynamic provisioning for Persistent Volume, remove storageClassName

```

vim services/servingservice/servingservice-pvc.yaml

```

```

apiVersion: v1

kind: PersistentVolumeClaim

metadata:

  name: serving-service-claim

spec:

  accessModes:

    - ReadWriteOnce

  resources:

    requests:

      storage: 500Mi

```

11. Serving service runs multiple Pods, we must assign all Pod instances to the same Kubernetes node, to make sure all instances can access Persistent Volume. Read more - [Assigning Pods to Nodes](https://kubernetes.io/docs/concepts/scheduling-eviction/assign-pod-node/)

```

kubectl get nodes

```

```

kubectl label nodes  skipper=serving

```

```

kubectl get nodes --show-labels

```

12. Add nodeSelector to servingservice-pod.yaml

```

nodeSelector:

    skipper: serving

```

13. Setup Kubernetes services:

```

./kubectl-setup.sh

```

14. Skipper API endpoint URL

```

http:///api/v1/skipper/tasks/docs

```

Check Load Balancer IP:

![LoadBalancer](https://github.com/katanaml/katana-skipper/blob/master/gke-loadbalancer.png)

## License

Licensed under the Apache License, Version 2.0. Copyright 2020-2021 Katana ML, Andrej Baranovskij. [Copy of the license](https://github.com/katanaml/katana-pipeline/blob/master/LICENSE).