While we are on the topic of image pulling optimizations, following granular image layers level optimizations are being worked on as part of Kubernetes Scheduling SIG future release feature:  

1)      Layer tracking system on each machine/node discovers the layer information from local file-system and reports this information to workload scheduler (for example, K8S scheduler in our case).

2)      As part of the image resolution process for an incoming container workload, image request is decomposed to underlying layer requests.

3)      Scheduler uses these two pieces of information to perform workload scheduling in accordance to the layering information on the node.

Regards,

Deepak Vij