Automatic cross-node image synchronization using the embedded registry?

[mjrist]

When using the embedded registry is it possible to automatically pull images onto every node regardless of whether the image is currently needed on the node. For high availability, I want the containerd image store on every node to contain the complete set of all images. So when I import an image to one node I would like all additional nodes in the cluster to immediately pull that image.

Example:

I have a three node cluster (all master nodes) A, B, and C

I manually import image1 on node A. I schedule a deployment that uses image1. Let's say it gets scheduled on node A. The image is already on that node so nothing gets pulled.

A while later hardware failure occurs and node A goes down. The pod for the deployment gets rescheduled onto node B. The image was only on node A. Since it is no longer available the container cannot be created.

Instead if images were synchronized on import or periodically, the image would be available on every node.

[brandond]

When using the embedded registry is it possible to automatically pull images onto every node regardless of whether the image is currently needed on the node.

No. Spegel does not currently support anticipatory replication. Images are only pulled by containerd when requested - either by the kubelet, or via ctr/crictl CLI tools.

You might open a feature request: https://github.com/spegel-org/spegel

[phillebaba]

I think it is in interesting feature to discuss for Spegel or a subproject in spegel-org, but as @brandond lets move such discussions to the Spegel repo.