> The chunk boundaries should be determined deterministically from local properties of the data. Use a rolling checksum over some small window and split the file it it hits a special value (0). This is what the rsyncable patch to zlib does.
> We already have the rsync algorithm which can scan through a file and efficiently tell which existing chunks match (portions of) it, using a rolling checksum. (Here's a refresher: http://samba.anu.edu.au/rsync/tech_report/node2.html ). Why not treat the 'chunk' as the fundamental unit, and compose files from chunks?
Prolly Trees are Merkle-fied B-trees, essentially.
I am working on related things[r], using Merkle-fied LSM trees. Ink&Switch do things that very closely resemble Merklefied LSM[h], although they are not exactly LSM. I would not be surprised if someone else is doing something similar in parallel. The tricks are very similar to Prollies, but LSM instead of B-trees.
That reminds me my younger years when I "invented" the Causal Tree[c] data structure. It was later reinvented as RGA (Replicated Growable Array [a]), Timestamped Insertion Tree and, I believe, YATA. All seem to be variations of a very very old revision control data structure named "weave"[w].
Recently I improved CT to the degree that warranted a new algorithm name (DISCONT [d]). Fundamentally the same, but much cheaper. Probably, we should see all these "inventions" as improvements. All the Computer Science basics seem to have been invented in the 70s, 80s the latest.
This is the first time I hear about librdx and Chotki, very cool projects. I skimmed over both projects but I haven't seen much written about conflict resolution. Does it mean that last write wins based on version vectors?
This article does not mention Jumbostore (Kave Eshghi, Mark Lillibridge, Lawrence Wilcock, Guillaume Belrose, and Rycharde Hawkes) which used content defined chunking recursively on the chunk list of a content defined chunked file in 2007. This is exactly what a Prolly Tree is.
The reason I thought a new name was warranted is that a prolly tree stores structured data (a sorted set of k/v pairs, like a b-tree), not blob data. And it has the same interface and utility as a b-tree.
Is it a huge difference? No. A pretty minor adaptation of an existing idea. But still different enough to warrant a different name IMO.
How do people find specialised data structure that they aren't already aware of? Stumbling across random blog posts and reading the odd book on data structures can't be the optimal way.
Is there a way to search for a structure by properties? E.g. O(1) lookups, O(log(n)) inserts or better, navigates like a tree (just making this up), etc?
Haha, this is funny. I've been obsessed with rolling-hash based chunking since I read about it in the dat paper. I didn't realize there was a tree version, but it is a natural extension.
I have a related cryptosystem that I came up with, but is so obvious I'm sure someone else has invented it first. The idea is to back up a file like so: first, do a rolling-hash based chunking, then encrypt each chunk where the key is the hash of that chunk. Then, upload the chunks to the server, along with a file (encrypted by your personal key) that contains the information needed to decrypt each chunk and reassemble them. If multiple users used this strategy, any files they have in common would result in the same chunks being uploaded. This would let the server provider deduplicate those files (saving space), without giving the server provider the ability to read the files. (Unless they already know exactly which file they're looking for, and just want to test whether you're storing it.)
Tangent: why is it that downloading a large file is such a bad experience on the internet? If you lose internet halfway through, the connection is closed and you're just screwed. I don't think it should be a requirement, but it would be nice if there was some protocol understood by browsers and web servers that would be able to break-up and re-assemble a download request into a prolly tree, so I could pick up downloading where I left off, or only download what changed since the last time I downloaded something.
"Tangent: why is it that downloading a large file is such a bad experience on the internet?"
This comment could only come from someone who never downloaded large files from the internet in the 1990s.
Feels like heaven to me downloading today.
Watching video from YouTube, Facebook, etc., if accessed via those websites running their Javascripts, usually uses the Range header. Some people refer to the "breeak up and re-assembly" as "progressive download".
Adding tangent to tangent, I recently experienced an unexpected modern counterpart of a 1990s large download: deleting about 120K emails from a GMail folder, then purging them for real by "emptying" the GMail "trash bin".
The first phase was severely asynchronous, with a popup mentioning "the next few minutes", which turned out to be hours. Manually refreshing the page showed a cringeworthy deletion rate of about 500 messages per minute.
But at least it worked; the second phase was more special, with plenty of arbitrary stopping and outright lies. After repeated purging attempts I finally got an empty bin achievement page on my phone but I found over 50K messages in the trash on my computer the next day, where every attempt to empty the trash showed a very slow progress dialog that reported completion but actually deleted only about 4K messages.
I don't expect many JavaScript card castles of the complexity of GMail message handling to be tested on large jobs; at least old FTP and web servers were designed with high load and large files in mind.
> If you lose internet halfway through, the connection is closed and you're just screwed. [...] it would be nice if there was some protocol understood by browsers and web servers
HTTP Range Requests solve this without any clever logic, if mutually supported.
> An HTTP Range request asks the server to send parts of a resource back to a client. Range requests are useful for various clients, including media players that support random access, data tools that require only part of a large file, and download managers that let users pause and resume a download.
AES-GCM-SIV[1] does something similar to your per chunk derived key, except that AES-GCM-SIV expects the key to be user-provided, and the IV is synthetic - hence Synthetic IV mode.
What's your threat model? This has "interesting"[3] properties. For example, given a file, the provider can figure out who has the file. Or, given a file, an arbitrary user can figure out if some other user already has the file. Users may even be able to "teleport" files to each other, like the infamous Dropbox Dropship[2].
I suspect why no one has tried this is many-fold: (1) Most providers want to store plaintext. Those few providers who don't want to store plaintext, whether for secrecy or deniability reasons, also don't want to store anything else correlatable, either. (2) Space is cheap. (3) Providers like being able to charge for space. Since providers sell space at a markup, they almost want you to use more space, not less.
I think the cost of processing stuff that way would far exceed the cost of downloading the entire file again. You can already resume downloads from a byte offset if the server supports it, and that probably covers 99% of the cases where you would actually want to resume a download of a single file. Partial updates are rarely possible for large files anyway, as they are often compressed. If the host wants to make partial updates make sense then they could serve over rsync.
Anyone know if editing a prolly tree requires reconstructing the entire tree from the leaves again? All the examples I've ever seen in a wild reconstruct from the bottom up. Presumably, you can leave the untouched leaves intact, and the reconstruct parent nodes whose hashes have changed due to the changed leaves. I ended up doing an implementation of this, and wondered if it's of any interest or value to others?
I am confused by this question. Both noms and dolt, and presumably most other prolly tree implementations do what you propose. if they didn't, inserts would be terribly slow.
The opposite of probabilistic is not deterministic in this context. This is not about «drawing a random number», but rather that balancing is dependent on the input data. «With high probability» here means «majority of the possible input data leads to a balanced structure».
If it was not probabilistic then the balancing would be guaranteed in all cases. This typically means that it somehow stores balancing information somewhere so that it can detect when something is unbalanced and repair it. In this data structure we’re just hashing the content without really caring about the current balance and then it turns out that for most inputs it will be fine.
The specific behavior of the hash function (including the salt you chose). Choosing a different hash function (or a different salt) would result in a different breakdown into chunks.
In principle, if your data were specially crafted to exploit the specific hash function (and salt) you could get an aberrant case like 1 million entries in a single b-tree node or a million b-tree nodes with just one entry. But unless you intentionally exploit the hash function the chance of this is vanishingly small.
Here's a post to the git mailing list from Martin Uecker describing the same from 2005: https://lore.kernel.org/git/20050416173702.GA12605@macavity/ . From the tone of the email it sounds like he didn't consider the idea new at that point:
> The chunk boundaries should be determined deterministically from local properties of the data. Use a rolling checksum over some small window and split the file it it hits a special value (0). This is what the rsyncable patch to zlib does.
He calls it a merkle hash tree.
Edit: here's one that's one day earlier from C. Scott Ananian: https://lore.kernel.org/git/Pine.LNX.4.61.0504151232160.2763...
> We already have the rsync algorithm which can scan through a file and efficiently tell which existing chunks match (portions of) it, using a rolling checksum. (Here's a refresher: http://samba.anu.edu.au/rsync/tech_report/node2.html ). Why not treat the 'chunk' as the fundamental unit, and compose files from chunks?
reply