Getting Started

Based on the official Kubernetes quick start (source 9). Commands assume a working Kubernetes cluster and Helm 3.

Prerequisites

• A Kubernetes cluster with kubectl configured.
• Helm 3 installed.
• containerd as the node runtime, since the mirror configuration below targets containerd.

Install

Add the chart repository and update:

helm repo add dragonfly https://dragonflyoss.github.io/helm-charts/
helm repo update

A first working setup

The chart deploys the full set of roles: manager, scheduler, seed peer, and the dfdaemon client as a DaemonSet.

1. Install Dragonfly into its own namespace.

   helm install --create-namespace --namespace dragonfly-system dragonfly dragonfly/dragonfly

2. Point containerd at the Dragonfly proxy as a registry mirror so image pulls flow through Dragonfly. The chart documents the mirror configuration; apply it on each node and restart containerd. See the quick start for the exact config.toml snippet (source 9).

3. Pull an image as usual. The pull now goes through the Dragonfly peer, which fetches pieces from peers or seed peers and falls back to the origin only when needed.

Verify it works

Confirm the pods are running:

kubectl get pods --namespace dragonfly-system

You should see the manager, scheduler, seed-peer, and dfdaemon pods in Running state. To confirm pulls are being accelerated, pull the same image on a second node and check the dfdaemon logs for piece downloads served from peers rather than the origin.

Where to go next

For production concerns such as high availability, the database backing the manager, TLS between roles, and tuning the scheduler, see the official Dragonfly documentation (source 9). For AI model distribution with hf:// and modelscope:// sources, see the CNCF write-up (source 8).