Pull Once, Scale Everywhere: Fixing Image Pull Bottlenecks With Dragonfly
summary
Dragonfly replaces thousands of concurrent registry pulls with a cluster-local peer-to-peer distribution using manager, scheduler, seed peer and per-node daemon, cutting startup latency and egress costs; demo shows a first pull then subsequent node pulls dropping from initial ~6–20s to ~12s as chunks propagate. It supports containerd/CRI-O/Podman, stores metadata in MySQL/Redis, offers image preheating and garbage collection, routes around overloaded peers via the seed peer pulls once pattern and falls back to the original registry if components fail.
AI-written summary · about
chapters (8)
- 0:00Kubernetes image pull challenges
- 2:29Image pull bottlenecks slow scaling
- 5:18Peer-to-peer architecture benefits
- 7:27Dragonfly’s core components
- 11:26Cost reduction with Dragonfly
- 14:18Preheating images for scaling
- 17:55Live demo in Kubernetes
- 24:59Q&A on seed peer behavior
related talks
29mAccelerates Image Distribution In Multi-Cluster With Dragonfly
34mDeep Dive: Dragonfly
27mDragonfly: Make Image Distribution Efficiently and Safely in Cloud Native
37mIntro: Dragonfly
34mFast Image Pulls Using IPFS And Opportunistic Caching
34mThree Approaches to Speed up Image Distribution in Cloud Native Era
31mThe New Evolution of Image Distribution System Dragonfly
34mThree Approaches to Speed up Image Distribution in Cloud Native Era