Skip to main content
Feature Engineering is deployed automatically in LanceDB EnterpriseIn self-managed environments, Geneva can be installed into existing Kubernetes clusters using Helm. Please contact LanceDB for access to the Helm Chart and related resources.

​
Pre-requisites

  • An existing Kubernetes cluster
  • An existing node pool(s) for Geneva workloads. By default, Geneva uses node selector {"geneva.lancedb.com/ray-head": "true"} for Ray head nodes, and {"geneva.lancedb.com/ray-worker-cpu": "true"} and {"geneva.lancedb.com/ray-worker-gpu": "true"} for Ray CPU worker and Ray GPU worker nodes respectively. This can be overridden in the Geneva client.
  • Geneva Helm chart. Please contact LanceDB for access to the Helm Chart and related resources.
For more information on deploying the required cloud resources, see the manual deployment instructions.

​
Geneva Helm Chart

The Helm chart includes resources required for running Geneva in Kubernetes. It includes services, service accounts, RBAC roles, etc. that are used by the Geneva client to manage resources.

​
Install

  1. Authenticate with Kubernetes cluster, i.e. update kubeconfig
  2. Configure Helm chart values
In values.yaml, configure the service account, node selectors, and cloud resources, if applicable. 
geneva:
  # Object storage root URI
  rootUri:
    value: "s3://my-data-bucket"

  serviceAccount:
    # Service account for Geneva worker pods and services
    annotations:
      # Set per-CSP annotations to provide access to CSP resources, i.e.
      # eks.amazonaws.com/role-arn: arn:aws:iam::0123456789:role/geneva_service_role
      # iam.gke.io/gcp-service-account: geneva-service-account@my-project.iam.gserviceaccount.com

  gcp:
    # GCP service account email for the Geneva client.
    # It should have access to the GKS cluster and "roles/storage.objectUser"
    # permissions on the object storage bucket.
    # e.g., geneva-client-sa@project-id.iam.gserviceaccount.com
    clientServiceAccount: ""

  aws:
    # AWS IAM role ARN to be assumed by the Geneva client.
    # This role should have an access entry to the cluster with username matching the role ARN.
    # It should also have r/w access to the object storage bucket.
    # e.g., arn:aws:iam::123456789012:role/geneva-client-role
    clientRoleArn: ""

  azure:
    # Azure managed identity client ID for the Geneva client.
    # This identity should have a federated credential for the LanceDB namespace
    # and Storage Blob Data Contributor role on the storage account.
    clientPrincipalId: ""
  1. Install kuberay operator
export NAMESPACE=lancedb

helm repo add kuberay https://ray-project.github.io/kuberay-helm/
helm repo update
helm install kuberay-operator kuberay/kuberay-operator -n $NAMESPACE --create-namespace
  1. Install NVIDIA device plugin (if using GPU nodes)
For GPU support, the NVIDIA device plugin must be installed in your EKS cluster: 
curl https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.17.0/deployments/static/nvidia-device-plugin.yml > nvidia-device-plugin.yml
kubectl apply -f nvidia-device-plugin.yml
  1. Install Geneva Helm chart
helm install geneva ./geneva -n $NAMESPACE --create-namespace

​
Default cluster and manifest

In LanceDB Enterprise, backfill and refresh jobs run on a default cluster (the compute pool jobs run on) and a default manifest (the Python dependency environment — image and packages). Configuring these in the LanceDB Enterprise chart lets jobs run out of the box without per-job configuration. They are set under geneva.defaults in the chart’s values.yaml: 
geneva:
  defaults:
    cluster:
      cluster_type: external_ray
      name: deployment-default
      ray_address: "ray://raycluster-kuberay-head-svc.lancedb.svc.cluster.local:10001"
    manifest:
      name: deployment-default
      pip: [geneva, pyarrow, lancedb, pylance]
      head_image: rayproject/ray:2.54.0-py312
      worker_image: rayproject/ray:2.54.0-py312
      skip_site_packages: true
If no default is configured, jobs must specify a manifest explicitly. Individual transforms can override the default manifest by pinning one with @udf / @chunker / @udtf (see Advanced Job Configuration); to override the cluster at runtime, use an Advanced Execution Context.

​
Providing a Ray cluster

The LanceDB Helm chart can be configured to deploy a static KubeRay cluster, provision KubeRay clusters on demand per job, or use an existing Ray cluster.

​
Use default LanceDB Enterprise Ray cluster (default)

By default, LanceDB Enterprise will use a shared, statically provisioned Ray cluster for job execution. This can be enabled in the Helm chart by setting the following values. 
raycluster:
    enabled: true

global:
    rayclusterUri: "ray://raycluster-kuberay-head-svc.lancedb.svc.cluster.local:10001"
Configuration for the Ray cluster can be specified by modifying raycluster.yaml Helm values.

​
Provision KubeRay clusters on demand

Set global.rayclusterUri to an empty value to provision ephemeral KubeRay clusters on-demand for each execution job. The default KubeRay cluster configuration is specified in geneva.defaults.cluster, i.e. 
geneva:
  defaults:
    cluster:
      cluster_type: kuberay
      name: deployment-default
      kuberay:
        namespace: lancedb
        config_method: IN_CLUSTER
        head_group:
          service_account: geneva-service-account
          num_cpus: 2
          memory: 8Gi
          image: rayproject/ray:2.54.0-py312
        worker_groups:
          - name: cpu
            service_account: geneva-service-account
            num_cpus: 4
            memory: 8Gi
            replicas: 2
            min_replicas: 0
            max_replicas: 4
            idle_timeout_seconds: 60
            node_selector:
              geneva.lancedb.com/ray-worker-cpu: "true"
            image: rayproject/ray:2.54.0-py312

​
Use an external Ray cluster

Self-managed enterprise customers can bring an existing Ray cluster to run Geneva jobs. Simply set the rayclusterUri property in the Helm chart to a Ray address that can be accessed from the LanceDB Enterprise deployment. 
global:
  rayclusterUri: "ray://my-ray-cluster.my-ns.svc.cluster.local:10001"