Skip to main content
H2O ships an official Docker image and a Helm chart for Kubernetes. Docker is suitable for single-node local or CI use; Kubernetes is the recommended path for multi-node production deployments.

​
Running H2O in Docker

​
Official image

The H2O Docker image is based on Ubuntu 24.04 and includes:
  • OpenJDK 8
  • Python 3.11 with h2o, scikit-learn, pandas, numpy, and related packages
  • The latest stable h2o.jar
Exposed ports:
  • 54321 — H2O Flow web UI and REST API
  • 54322 — internal node-to-node communication

​
Quick start

docker run -ti -p 54321:54321 h2oai/h2o-open-source-k8s /bin/bash
Inside the container, start H2O: 
java -jar /opt/h2o.jar
Then open http://localhost:54321 in your browser to access H2O Flow.

​
Building a custom image

Use the Dockerfile from the H2O repository as a starting point: 
# Clone and build
git clone https://github.com/h2oai/h2o-3.git
cd h2o-3
docker build -t my-h2o:latest .
The official Dockerfile:
  1. Installs Java 8 and Python 3.11 on Ubuntu 24.04
  2. Downloads the latest stable h2o.jar from H2O’s S3 repository
  3. Installs the H2O Python wheel
  4. Exposes ports 54321 and 54322
For reproducibility in production, pin to a specific H2O version by setting the image tag rather than using the latest tag.

​
Kubernetes deployment via Helm

The h2o-helm chart deploys a multi-node H2O cluster as a Kubernetes StatefulSet. Nodes discover each other automatically through the Kubernetes DNS service.

​
Prerequisites

  • Helm 3.x
  • A Kubernetes cluster with sufficient CPU and memory resources
  • kubectl configured to point at your cluster

​
Installing the chart

helm install h2o-cluster ./h2o-helm \
    --set h2o.nodeCount=3 \
    --set image.tag=3.46.0.1 \
    --set resources.memory=4Gi \
    --set resources.cpu=2

​
Minimal values.yaml

values.yaml
image:
  name: h2oai/h2o-open-source-k8s
  tag: "3.46.0.1"

h2o:
  nodeCount: 3          # Number of H2O nodes in the cluster
  memoryPercentage: 50  # Percentage of container memory allocated to H2O JVM heap
  lookupTimeout: 180    # Seconds to wait for all pods to join the cluster

resources:
  cpu: 2
  memory: 4Gi
Apply via Helm: 
helm install h2o-cluster ./h2o-helm -f values.yaml

​
Example with ingress

example.yaml
ingress:
  enabled: true
  annotations: {}
  hosts:
    - host: h2o.example.com
      paths: ["/"]
  tls: []

h2o:
  nodeCount: 3

image:
  tag: 3.30.1.2

​
How cluster formation works in Kubernetes

H2O uses a StatefulSet with podManagementPolicy: Parallel so all pods start simultaneously. Each node discovers its peers via the Kubernetes headless service DNS name: 
<release-name>.<namespace>.svc.cluster.local
The relevant environment variables injected into each pod:
VariableDescription
H2O_KUBERNETES_SERVICE_DNSHeadless service DNS for peer discovery
H2O_NODE_LOOKUP_TIMEOUTSeconds to wait for expected peers (default 180)
H2O_NODE_EXPECTED_COUNTTotal number of nodes to wait for
H2O_KUBERNETES_API_PORTPort for the Kubernetes readiness probe (default 8080)
A readiness probe at /kubernetes/isLeaderNode on port 8080 gates traffic until the leader node is elected and the cluster is fully formed.

​
Scaling nodes

H2O clusters cannot be horizontally scaled after formation — all nodes must be present from the start and remain up for the duration of the cluster’s lifetime. Rescaling requires restarting the entire StatefulSet.
To change the cluster size, update the Helm release: 
helm upgrade h2o-cluster ./h2o-helm --set h2o.nodeCount=5
This triggers a full StatefulSet rollout. Any in-progress training jobs will be lost. For reproducible resource allocation, resources.cpu and resources.memory are set as both requests and limits — Kubernetes will always give H2O exactly the resources specified.

​
Accessing the H2O UI and API

By default the chart creates a ClusterIP service on port 80. To expose H2O externally: Load balancer (for cloud environments): 
loadbalancer:
  enabled: true
Port forwarding (for local development): 
kubectl port-forward svc/h2o-cluster 54321:80
# Then open http://localhost:54321
Ingress (for production HTTPS exposure): 
ingress:
  enabled: true
  hosts:
    - host: h2o.example.com
      paths: ["/"]
  tls:
    - secretName: h2o-tls
      hosts:
        - h2o.example.com

​
Helm chart reference

The chart source lives at h2o-helm/ in the H2O-3 repository.
FilePurpose
Chart.yamlChart metadata and version
values.yamlDefault configuration values
templates/statefulset.yamlH2O node StatefulSet definition
templates/service.yamlHeadless service for peer discovery
templates/ingress.yamlOptional ingress resource
templates/loadbalancer.yamlOptional load balancer service

Build docs developers (and LLMs) love

Get started for freeTalk to us