This came up in a discussion elsewhere, just tossing it up on the list if anyone's interested:
> Doesn't this suggest that [Docker sharing] should be independent of the way > images are constructed? This could probably be implemented even today > with overlayfs, by keeping a hash file farm and hard-linking that into > the immutable base image, similar to what OSTree does. The overlayfs backend right now does use hardlinks between base and derived, relying on overlayfs copy-up to avoid "hardlink writethrough" - but What the above is suggesting is the "hidden ostree-izing" model. It's quite possible. This would involve something like the overlayfs backend being modified to read/write to the host ostree repo. It'd be a bit messy as then docker's overlayfs backend would depend on an OSTree setup - so it'd probably have to be optional. But a neat feature of this model is it would enable sharing with the *host*, not just between containers. For example if you have the exact same /usr/lib64/libc-2.17.so or /usr/bin/bash objects in both, which would be quite common. It would also enable dynamic sharing between independent containers that have the same content - this would happen with e.g. new base images generated from kickstarts instead of deriving from $distro base.
