Deploy with Helm

BenchGen ships as one Helm chart (benchgen) that deploys the whole platform: the core benchmarking stack (Django, Prefect, LiteLLM, MinIO, compute worker) plus the required agentspace bundle (chat UI, AI-bot agent, and their backing services). The chart carries working defaults for everything structural. Your values file only needs the environment-specific inputs: hostnames, secrets, and per-service credentials. See the Helm values reference for every parameter.

Prerequisites

Requirement	Notes
Kubernetes cluster	Tested on k3s; any conformant cluster works
Ingress controller	Traefik or NGINX - set `ingress.provider` to match
cert-manager	With a `ClusterIssuer` named `letsencrypt-prod` for TLS certificates
Pull secret	A `dockerhub-creds` secret in the target namespace for the private `betedo/*` images
DNS	Every hostname in your values file must point at the cluster (required for routing and HTTP-01 certificate issuance)
GPU server (optional)	Ray with vLLM for local model inference - can run remotely; see `config.RAY_URL`

Create the pull secret before installing:

kubectl create namespace benchgen
kubectl -n benchgen create secret docker-registry dockerhub-creds \
  --docker-username=<user> --docker-password=<token>

For air-gapped environments the images are served from the platform registry instead of the public internet - see Air-gapped environments on the DT Edge Platform page.

Install

Start from the minimal values file (values-minimal.yaml in the chart) and replace every CHANGE-ME placeholder and hostname:

helm upgrade --install benchgen ./benchgen-<version>.tgz \
  -n benchgen --create-namespace \
  -f my-values.yaml

The release includes post-install hooks that bootstrap the platform automatically:

Database migrations run before the app starts.
job-chatui-admin-init creates the chat UI admin account and applies instance settings (open sign-ups, default role and permissions, API connections) through the admin API.
job-aibot-admin-init seeds the agentspace admin and the client account through the AI-bot agent API.

Both hooks are idempotent - they detect existing accounts and skip creation on re-runs.

Ray is the only optional dependency. Keep ray.enabled: false and point config.RAY_URL / config.RAY_MODEL_API at a remote GPU server, or enable it in-cluster on a GPU node.

Upgrade

Upgrades are plain Helm upgrades with a newer chart version:

helm upgrade benchgen ./benchgen-<new-version>.tgz -n benchgen -f my-values.yaml

Only pods whose rendered spec changed are restarted. Application data lives on PersistentVolumeClaims and survives upgrades.

StatefulSet volumes (shared PostgreSQL, RabbitMQ, Redis, vector DB) survive even a helm uninstall. PVCs defined by the chart itself (agent databases, chat UI, Grafana/Loki/Prefect) are deleted with the release - back them up before removing it.

Verify the deployment

# All pods should reach Running/Completed
kubectl -n benchgen get pods

# Hook jobs should show Completed
kubectl -n benchgen get jobs

# Certificates should become Ready
kubectl -n benchgen get certificates

Then open https://<ingress.hosts.main> and sign in with the superadmin account (superadmin / secrets.SUPERADMIN_PASSWORD).

Get started

Agents

Eval

Train

Deploy Benchgen Platform

Prerequisites

Install

Upgrade

Verify the deployment

​Prerequisites

​Install

​Upgrade

​Verify the deployment

Prerequisites

Install

Upgrade

Verify the deployment