Preheated images do not improve image pull time, and all-seed-peers scope still downloads layers from Harbor Registry.

[swatiskadam578]

Description

I have deployed Dragonfly using the Helm chart and configured Harbor as the image registry.

After preheating an image, I observed the following behavior:

1. Preheat Scope: all-peers

   • Successfully preheated the image with scope all-peers.
   • Deployed a workload on the same Kubernetes cluster.
   • The image pull time was almost identical to pulling the image without Dragonfly.
   • I expected the image to be served from Dragonfly peers, resulting in a noticeable reduction in pull time.

2. Preheat Scope: all-seed-peers

   • Successfully preheated the image with scope all-seed-peers.
   • Deployed the same workload.
   • I observed requests still reaching the Harbor registry while the image was being pulled.
   • I expected the image to be served primarily from Dragonfly seed peers without requiring Harbor for the image layers (unless a cache miss occurred).

Expected Behavior

• Images preheated with all-peers should reduce image pull time compared to pulling directly from Harbor.
• Images preheated with all-seed-peers should be served from Dragonfly peers/seed peers, with minimal or no layer downloads from Harbor when the image has already been preheated.

Actual Behavior

• No improvement in image pull time after preheating with all-peers.
• Harbor registry is still accessed during image pulls after preheating with all-seed-peers.

Environment

• Dragonfly: Helm chart deployment
• Kubernetes:
• Dragonfly Helm Chart Version:
• Harbor Version:
• Container Runtime: <containerd/docker>
• Image Size: <e.g. 9 GB>

Steps to Reproduce

1. Deploy Dragonfly using the Helm chart.
2. Configure Harbor to use Dragonfly.
3. Preheat an image using scope all-peers.
4. Deploy a pod using the preheated image.
5. Observe the image pull time.
6. Repeat using all-seed-peers.
7. Observe that Harbor registry is still contacted during image pull.

Additional Information

Could you please clarify:

• Is this the expected behavior?
• Are there any additional Helm values or configuration required for preheated images to be served entirely through Dragonfly?
• How can I verify that image layers are actually being served from Dragonfly instead of Harbor?