Large-scale Kubernetes deployments suffer from the 'thundering herd' problem when hundreds of nodes simultaneously pull identical container images from a central registry, causing network saturation and ImagePullBackOff errors. CNCF-graduated project Dragonfly solves this by transforming Kubernetes clusters into P2P distribution meshes. Its three components — Manager, Seed Peer, and Dfdaemon DaemonSet — intercept containerd pull requests transparently and distribute image blocks node-to-node, reducing central registry load by up to 99%. Deployment is done via Helm chart with a containerd mirror config change. Beyond containers, Dragonfly now supports AI model distribution (Hugging Face, ModelScope), enabling P2P sharing of 100GB+ LLM weight files across GPU clusters over NVLink or InfiniBand.
