Session

Every Node Pulls Alone: Fixing Container Image Distribution with CNCF Dragonfly

Every time a Kubernetes node starts, it downloads every container image it needs from the registry, alone. With autoscaling and spot capacity creating thousands of fresh nodes a day, that means slow cold starts, a growing registry bill, and a single point of failure for every scale-up.

CNCF Dragonfly, a graduated project since January 2026, fixes this with peer-to-peer distribution: nodes share image layers with each other, and the registry serves each layer about once. We rolled it out across 30+ EKS clusters. On day one it looked great. Months later we found that only 1.4% of pulls on one of our biggest clusters actually went through it, while every dashboard stayed green.

This talk explains how image pulls work, how Dragonfly changes them, and the eleven things that broke in our rollout, from a boot race to a cache that pulled its own dependencies through itself. You'll leave able to measure how much of your image traffic really goes through your cache, and to gate new nodes until it's ready.

Wendell Beckwith

Redfin, Technical Lead, Cloud Orchestration

Actions

Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.

Jump to top