2018-11-08
I don’t like the way Docker stores and distributes images, for several reasons:
Other container systems do better on this front but don’t make much attempt to solve the image distribution problem.
What follows is a basic sketch of a container image format with efficient incremental distribution and none of the properties I complained about above.
Images should be SquashFS. That’s it. This is a solved problem. You don’t need to unpack SquashFS; you just mount it. It’s read-only.
Configuration (environment, entrypoint, etc.) could live inside the image. Perhaps there could be two top-level nodes:
config.json
, containing whatever settings;root/*
, containing the full image filesystem.Overlay. Mount the SquashFS read-only, then mount a read-write overlay on top. The container can be started using this overlay as the filesystem root.
Since SquashFS can be mounted directly, all you need to do is download a file and run it. However, downloading a marginally-changed new revision of an image results in a full re-download. We can solve this by diffing the two versions and generating a binary diff.
In the past, I’ve used xdelta3 for this. bsdiff is another option. However, both of these are slow: far to slow to generate on-the-fly in response to a client request.
A more timely way to do differential distribution would be to read both SquashFS files and generate a list of files changed between them, generating a file containing all of the changed files. Moved could be accounted for by taking a sum of the contents of each file as well. An old library I wrote—treediff—could be a starting point for this.
With a properly SquashFS-aware diff library, it should be feasible to have a repository server respond to requests for an image diff given a target image and a reference image, already possessed, and generate the diffs inline (ideally with caching).