Would be nice to see at least a high-level overview of how it works under the hood. Is it doing anything interesting with reusable layers? Actually, thinking further that might be a moot point; I feel like running as a remote proxy loses some fun optimization chances. I could imagine a world where you install a package from the host's pip/uv, and then add it to a container image, and both of those are actually the same thing on disk. (Granted, that's likely harder to implement)
It's basically a single layer image with a multi-platform image index per tag(version) on top.
No smarts added, performance wise it's probably worse than running your own dedicated pypi index or any purpose-built system, my main goal is a private index for if you don't have access to one already.
That's fair. I'm used to thinking of OCI images as being a particular optimization (there's a reason we don't ship containers around ass rootfs tarballs), but for your usecase it may well be a premature/pointless optimization.
There are still growing pains, but https://github.com/osbuild/bootc-image-builder exists and is likely to become exactly that in the general case (as it already is for the redhat family).
Oh those size limits are pushed plenty by AI images, no worries. I recently had a good laugh when I found a docker image that was 2 - 3 times as big as the OS partition of a lot of our smaller servers.
And our OS image build order would reuse layers better than those.