Kuaishou Container Cloud Platform Enhances Image Distribution with Dragonfly and Nydus

Challenge: Kuaishou's container cloud platform needed to provide ultra‑large‑scale, elastic, stable, and efficient image distribution for its rapidly growing and diverse services, with image distribution stability and efficiency being the most difficult problems.

Solution: The Kuaishou container cloud engineering team partnered with Alibaba Cloud and Ant Group to adapt Dragonfly and its sub‑project Nydus within the Longxi community. Dragonfly and Nydus proved the most suitable, offering strong compatibility with existing systems, a smooth transition path, and significant efficiency gains.

Effect: After deployment, the P2P network built by Dragonfly reduced Harbor's network bandwidth pressure by more than 70% on average and over 80% at peak, saved more than 90% of image pull time, and cut POD service latency by about 50%. The system now supports a far larger number of concurrent image pulls, eliminating bottlenecks during DaemonSet deployments and large‑scale service updates.

Projects Used: Dragonfly ( https://github.com/dragonflyoss/Dragonfly2 ), Nydus ( https://github.com/dragonflyoss/image-service ), containerd ( https://github.com/containerd/containerd ), and Harbor ( https://github.com/goharbor/harbor ).

Key Metrics: Peak bandwidth relief >80%; image pull time saved >90%; POD instance service time saved >50%.

Impact: The solution now supports stability and performance for over 1 billion monthly users. Leaders from Kuaishou highlighted that Dragonfly effectively solved massive file distribution problems and that the combined technologies dramatically improve application delivery efficiency.

Implementation Details: Engineers switched the container engine from Docker to containerd for better integration with Dragonfly, configured Dragonfly's P2P proxy for image pulls, and adopted Nydus for lazy‑loading images. This enables containers to start within seconds, even for large images, without requiring pod eviction.

Conclusion: Dragonfly and Nydus together provide the optimal solution for Kuaishou's image distribution challenges, reducing deployment time, easing service updates, and reinforcing Kuaishou's commitment to advancing cloud‑native open‑source projects.