Deduplicating Binary Version Control with Dupver

I deal a lot with large-ish binary file in my work. Typically these are single-file database in the style of Sqlite3. These aren't ideal for putting in Git. Git doesn't scale well to large repository sizes. People always seem to try framing this as not a limitation, but it is. I've taken to using Boar, Restic and finally the Duplicacy deduplicating backup software packages. I think at some point someone had mentioned that it would be awesome to have something like Git but that replaces the packfile format with deduplicating storage. And so I got started scratching my own itch. You can find it on my GitHub here:

https://github.com/akbarnes/dupver

Dupver is a minimalist deduplicating version control system in Go based on the Restic chunking library. It is most similar to the binary version control system Boar:

https://bitbucket.org/mats_ekberg/boar/wiki/Home

Dupver does not track files, rather it stores snapshots more like a backup program. Rather than traverse directories itself, Dupver uses an (uncompressed) tar file as input. Not that only tar files are accepted as input as Dupver relies on the tar container to provide the list of files in the snapshot and store metadata such as file modification times and permissions Dupver uses a centralized repository to take advantage of deduplication between working directories. This means that dupver working directories can also be git repositories or subdirectories of git repositories. I mainly use it for version control of databases, but it can also be used for sampled data.

Art, 2020-10-24 07:41

Update 01/02/2021: Some Thoughts on the Future of Version Control

In response to a HN post, some thoughts about the future for version control. Full discussion is here: https://news.ycombinator.com/item?id=25535844

The state-of-the-art for backup is deduplicating software (Borg, Restic, Duplicacy). Gripes about Git's UI choices aside, Git was designed around human-readable text files and just doesn't do large binary files well. Sure, there's Git-LFS, but it sucks. The future of version control will:

Make use of deduplication to handle large binary files Natively supports remotes via cloud storage Doesn't keep state in the working directory so that projects can live in a Dropbox/OneDrive/iCloud folder without corrupting the repo Is truly cross-platform with minimal POSIX dependencies. I love Linux, but I'm a practicing engineer, and the reality is that engineering software is a market where traditional Windows desktop software still rules. Another thought I've been having for some time is if I could have gotten away with file level deduplication like Boar (or Git IIRC) does and drop compression. This would probably result in significant simplification, particularly for copying between repos. For most users this wouldn't impact disk space usage much as the bulk of files already have compression built in, and the trend seems to be increasingly to adopt compression in new file formats. This includes:

Audio/Image/Video files with (usually) lossy compression. This suprisingly (to me) also includes raster image editor file formats such as Paint.net's pdn, which wraps everything in a gzip stream. MS office documents structured as a hierarchy of zipped .xml files. More recently, this format also includes Matlab's .slx Simulink file format and .mlx notebook format. The gotcha to this is it's an 80% solution. There are still plenty of file formats that are uncompressed text, even newer ones such as JSON/YAML/TOML and a number of uncompressed binary file formats such as MessagePack, though most tend to be some sort of database such as the Geodatabase .gdb format which is based on Sqlite3 or PowerWorld's .pwb format. There is also the corner case of metadata in media files such as EXIF, which if modified would cause the whole file contents to be stored again. So I'm sticking with chunking for the time being.