Story

Hi folks,

I am working on a game preservation project, where the data set holds 10.4 TiB. It contains 1044 earlier versions of a single game in a multitude of different languages, architectures and stages of development. As you can guess, that means extreme redundancy.

The goals are: - bring the size down - retain good read speed (for further processing/reversing) - easy sharable format - lower end machines can use it

My choice fell on the BTRFS filesystem, since it provides advanced features for deduplication, which is not as resource hungry as ZFS. Once the data is processed, it no longer requires a lot of system resources.

In the first round of deduplication, I used "jdupes -rQL" (yes, I know what -Q does) to replace exact copies of files in different directories via hardlinks to minimize data and metadata. This got it down to roughly 874 GiB already, out of which 866 GiB are MPQ files. That's 99,08%... everything besides is a drop in the bucket.

For those uninitiated: this is an archive format. Representing it as a pseudo-code struct it looks something like this { header, files[], hash_table[], block_table[] } Compression exists, but it is applied to each file individually. This means the same file is compressed the same way in different MPQ archives, no matter the offset it happens to be in.

What is throwing a wrench into my plans of further data deduplication are the following points: - the order of files seems not to be deterministic when MPQ files were created (at least I picked that up somewhere) - altered order of elements (files added or removed at the start) causes shifts in file offsets

I thought for quite some time about this, and I think the smartest way forward is, that I manually hack apart the file into multiple extents at specific offsets. Thus the file would contain of an extent for: - the header - each file individually - the hash table - the block table It will increase the size for each file of course, because of wasted space at the end of the last block in each extent. But it allows for sharing whole extents between different archives (and extracted files of it), as long as the file within is content-wise the same, no matter the exact offset. The second round of deduplication will then be whole extents via duperemove, which should cut down the size dramatically once more.

This is where I am hanging right now: I don't know how to pull it off on a technical level. I already was crawling through documentation, googling, asking ChatGPT and fighting it's hallucinations, but so far I wasn't very successful in finding leads (probably need to perform some ioctl calls).

From what I imagine, there are probably two ways to do this: - rewrite the file with a new name in the intended extent layout, delete the original and rename the new one to take it's place - rewrite the extent layout of an already existing file, without bending over backwards like described above

I need is a reliable way to, without chances of the filesystem optimizing away my intended layout, while I write it. The best case scenario for a solution would be a call, which takes a file/inode and a list of offsets, and then reorganizes it into that extents. If something like this does not exist, neither through btrfs-progs, nor other third party applications, I would be up for writing a generic utility like described above. It would enable me to solve my problem, and others to write their own custom dedicated deduplicaton software for their specific scenario.

If YOU - can guide me into the right direction - give me hints how to solve this - tell me about the right btrfs communities where I can talk about it - brainstorm ideas I would be eternally grateful :) This is not a call for YOU to solve my problem, but for some guidance, so I can do it on my own.

I think that BTRFS is superb for deduplicated archives, and it can really shine, if you can give it a helping hand.