Mentioning the more portable desync is obligatory: https://github.com/folbricht/desync
When I have big 100MB binary files that i want to version, the changes are small (1MB in one place, and a few more KB in others). I also have a few multi GB SQLite databases I would like to version where this would help. (Changes are less than 5% of the data, but index pages throughout the file mean rsync-style partitions still transfer 50% or so of the file. To actually achieve storage efficiency, I store textual dumps, and they also fit better with borg/restic/casync than git or git-lfs
Would have been extra awesome if desync would be able to use a git repo as storage.
For text files with really small changes, this is comparable to a "diff" size, and is better than what a rolling hash would achieve. However, for text files with larger changes, and of course for binary files - a rolling hash is much more effective. Additionally, a rolling hash would easily reuse cross-file similarity, where as the current delta-finding code is likely to only find redundancy in the history of the same file (or similarly named in the same directory).