Skip to content

Latest commit

 

History

History
102 lines (81 loc) · 4.1 KB

README.md

File metadata and controls

102 lines (81 loc) · 4.1 KB
Raw

H2O Kubernetes Operator

Rust

A command line tool to ease the deployment (and undeployment) of H2O open-source machine learning platform H2O-3 to Kubernetes. Currently in beta, with stable basic deployment/undeployment features. Official H2O Kubernetes Docker images are used.

Download for Mac/Linux/Windows.

Usage

H2O Usage in console

Once the h2o-operator is running, an H2O resource can be used together with kubectlto deploy H2O into Kubernetes. Example h2o.yaml YAML:

apiVersion: h2o.ai/v1
kind: H2O
metadata:
  name: h2o-test
spec:
  nodes: 3
  version: "3.32.0.1"
  resources:
    cpu: 1
    memory: "512Mi"
    memoryPercentage: 90

After creating the resource by usingkubectl apply -f h2o.yaml, all the necessary H2O resources are created. Deletion is as simple as kubectl delete h2o h2o-test.

Deployment

Operator images are available on:

Or build from scratch:

docker build --build-arg H2O_VERSION=custom -t h2o-operator:custom -f docker/public/Dockerfile-h2o-release build/

Once the Dockerfile is built, make sure to push it to a proper repository reachable by the Kubernetes cluster (e.g. Docker Hub).

Before the actual image is deployed, the H2O CustomResourceDefinition must be created in the Kubernetes cluster. The definition is to be found in bundle/manifests/h2o.crd.yaml. Download it and do kubectl apply -f h2os.h2o.ai.crd.yaml. Such an operation requires user with the following permissions:

- apiGroups:
    - "apiextensions.k8s.io"
  resources:
    - customresourcedefinitions
  verbs:
    - create

Once the H2O CRD is deployed, the operator itself may also be deployed. A simple Deployment with exactly one instance of the pod with H2O operator inside will do. In case of failures, the pod is restarted and H2O-related events are handled.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: h2o-operator
  labels:
    app: h2o-operator
spec:
  replicas: 1
  selector:
    matchLabels:
      app: h2o-operator
  template:
    metadata:
      labels:
        app: h2o-operator
    spec:
      containers:
      - name: h2o-operator
        image: repository/h2o-operator:custom # Replace with a real docker image specification
        imagePullPolicy: Always # Set to IfNotPresent if the image with the very same tag never changes

The operator requires specific permissions to run, too. Make sure to use a Kubernetes User or create a dedicated ServiceAccount with rights listed in the ClusterRole definition file. This set of permissions is used to test the operator itself.

Building, testing and running

Refer to the contributing guide for detailed instructions on how to build and develop this project.

  • Development build : cargo build -p h2o-operator
  • Release build: cargo build -p h2o-operator --release
  • Development run: cargo run -p h2o-operator -- deploy --namespace default --kubeconfig /etc/rancher/k3s/k3s.yaml --cluster_size 3 -- version latest
  • Test: cargo test -p h2o-operator - please note many tests have prerequisites - running Kubernetes cluster and the KUBECONFIG variable set.

On start, the h2o-operator binary will look for KUBECONFIG variable. If such variable exists and points to a valid kubeconfig file with enough permissions to run the operator, it will try to detect or create the H2O custom resource inside the Kubernetes cluster and start serving requests. It can be deployed either directly into Kubernetes inside a Pod or ran externally.

Technical documentation

Technical documentation signpost is to be found in the documentation folder.