← Journal · · Note
Read Only Volumes from OCI artifacts in Kubernetes 1.31
The new k8s 1.31 will add an interesting feature: Read Only Volumes Based On OCI Artifacts
You'll be able to mount another container into a specific location inside a pod, which gives us the ability (from an ML standpoint) to 1. Pack and version model weights in containers and store them in a container registry 2. Do the same with configs (for example, for Triton)
Right now, for instance, S3 is often used to store those same model weights (training artifacts), but that doesn't version very well. With this new approach, maybe we'll start pushing training artifacts into a container registry and using them that way in k8s)
And if you throw in snapshotters, or at least convert to zstd, download and deployment speed for weights could get even better...