Training cluster page

Enterprise only

The On-prem training cluster page is available only on Enterprise organizations. See Plan permissions.

Connect your organization’s private training hardware to Arena. Org Admins on Enterprise orgs open On-prem training cluster from the profile menu to connect a cluster and manage resource classes.

On Enterprise orgs the page has two parts: Enterprise K8s clusters, then on-prem resource classes. You do not enable a resource provider by hand; connecting your first cluster does that for you.

Prerequisites

You connect a cluster you already run. Before you start, have:

  • Enterprise plan on the current organization

  • Org Admin role

  • A running Kubernetes cluster with kubectl access

  • An S3-compatible bucket the cluster can reach

  • A Kubernetes Secret in the agent namespace holding that bucket’s credentials

  • Arena CLI installed (pip install "agilerl[arena]") and arena login to run the generated install command

Steps

1. Connect your cluster

With no clusters registered, the page shows Connect existing cluster. The form asks for four values only:

Field

Example

Name

corp-k8s

Storage endpoint

https://s3.corp.example.com

Storage bucket

arena-data

Storage secret name

arena-storage

Create the credentials Secret in your cluster first if it does not exist:

kubectl create secret generic arena-storage \
  --from-literal=AWS_ACCESS_KEY_ID=<key> \
  --from-literal=AWS_SECRET_ACCESS_KEY=<secret>

Cluster names are unique per org. Reusing a name updates that cluster instead of creating a duplicate.

2. Install the agent

Register saves the cluster and immediately shows the install handoff. The generated arena on-prem cluster register … --install --no-write command is the primary option — copy it and run it where you have kubectl access. Prefer charts and values instead? Open Install with Helm instead in the same panel.

Once the operator polls Arena, the cluster row shows Connected. Degraded means heartbeats stopped.

Later changes live on the cluster row: Install (regenerate files), Issue token / Rotate token, Edit, Disable, and Remove. Optional routing and storage settings (prefix, ingress class, hostname template, domain, TLS Secret, Ray data storage class and PVC size, preprocessing class) sit under Advanced in Edit and when registering an additional cluster. Hostname template is hostname-only (inference-{deploymentId}); domain is corp.example.com. An empty hostname template with a domain set defaults to inference-{deploymentId}. Arena joins them as {expanded_hostname_template}.{domain} when both are set.

3. Add resource classes

Add classes below the cluster table. Fields are described in Resource classes. With one connected cluster, new classes bind to it automatically. Refresh reloads the list.

4. Open deployment config

Click Config on a class row to expand the deployment panel inline.

Classes bound to an enterprise cluster are provisioned by the cluster agent and have no per-class Config panel. The panel below applies to legacy per-class installs.

Deployment configuration panel

Each cluster uses one install type — Docker Swarm or Helm on Kubernetes, not both. Panel order matches the UI:

  1. Arena CLI (recommended) — copy arena on-prem install and teardown commands for this class Name. Requires the CLI installed (pip install "agilerl[arena]"), arena login, and SSH to cluster hosts (Swarm only). The class already exists when you open Config from the table; the command installs workers for that class.

  2. Docker Swarm / Helm — toggle which template set and install command are shown.

  3. Download setup (.tar) — same bundle as the CLI; extract on a bastion, cd arena-train, run ./setup.sh or follow the README tab.

  4. File tabs below.

Tab (Swarm)

Content

README

Swarm install instructions (rendered markdown in the UI)

tun0.conf

WireGuard peer config

arena-stack.yaml

Stack definition

Tab (Helm)

Content

README

Helm install instructions

values.yaml

Chart values

head deployment

Head pod template

worker template

Worker pod template

Markdown tabs render in-page; YAML and conf files show as monospace text. Empty content shows as an em dash.

Install on hardware that matches the class CPU/GPU/memory and Number of nodes. Both Helm and Swarm bundles mount /opt/data on the Ray head and workers by default; multi-worker pools need shared NFS configured in values.yaml (Helm) or arena-stack.yaml (Swarm) — see Shared storage. Full install walkthrough: Install a cluster. Command reference: On-prem CLI.

Deleting a class

Delete opens a confirmation dialog naming the class. Deletion cannot be undone and removes the class from Arena; clean up the matching stack on your side separately.

Access

  • On-prem training cluster appears in the profile menu only for org Admins on Enterprise orgs.

  • Members on Enterprise orgs do not see the menu item. They can still pick on-prem classes on Resources once an Admin has configured them.

  • Registering, installing, and editing clusters requires an org Admin on an Enterprise org.