Edit: "The solution is to disaggregate the container launcher (runc for Docker, runsc for gVisor) from the container image delivery" is exactly what I've done! I've not built a lazy FUSE on top of it (yet! except for cache mounts in BuildKit), but it's on my TODO list. I guess I'm mainly curious what stops bytes from being shared in your case.

Docker is...quite slow with large images. I've built a registry+pull client+buildkit builder to make it better. It splits apart layers, allowing for files to be shared between related images. In a robotics context, it can make pulls 10x faster. And in a cloud context, the format allows for pulling an image in 15 or 20 seconds instead of 60, without having to do a FUSE w/lazy pulling. Builds are faster, I store 7x less data due to better deduplication, I can run security scans faster due to not having to unpack tarball layers, etc, etc. I want to be the default registry for all ML related work, in the future.

    After his release, Juettner briefly achieved celebrity status. His notoriety became so widespread that Hollywood adapted the story into the 1950 film Mister 880, directed by Edmund Goulding. Eventually, Juettner made more money from the release of Mister 880 than he had made by counterfeiting.

The documentation implies the last step is optional https://developer.apple.com/documentation/security/customizi... but it might be inaccurate