[astennumero]

What algorithm does the application use to figure out if two files are identical? There's a lot of interesting algorithms out there. Hashes, bit by bit comparison etc. But these techniques have their own disadvantages. What is the best way to do this for a large amount of files?

[borland]

I don't know exactly what Siracusa is doing here, but I can take an educated guess:

For each candidate file, you need some "key" that you can use to check if another candidate file is the same. There can be millions of files so the key needs to be small and quick to generate, but at the same time we don't want any false positives.

The obvious answer today is a SHA256 hash of the file's contents; It's very fast, not too large (32 bytes) and the odds of a false positive/collision are low enough that the world will end before you ever encounter one. SHA256 is the de-facto standard for this kind of thing and I'd be very surprised if he'd done anything else.

[MBCook]

You can start with the size, which is probably really unique. That would likely cut down the search space fast.

At that point maybe it’s better to just compare byte by byte? You’ll have to read the whole file to generate the hash and if you just compare the bytes there is no chance of hash collision no matter how small.

Plus if you find a difference in bytes 1290 you can just stop there instead of reading the whole thing to finish the hash.

I don’t think John has said exactly how on ATP (his podcast with Marco and Casey), but knowing him as a longtime listener/reader he’s being very careful. And I think he’s said that on the podcast too.

[jonhohle]

To make dedup[0] fast, I use a tree with device id, size, first byte, last byte, and finally SHA-256. Each of those is only used if there is a collision to avoid as many reads as possible. dedup doesn’t do a full file compare, because if you’ve found a file with the same size, first and last bytes, and SHA-256 you’ve also probably won the lottery several times over and can afford data recovery.

This is the default for ZFS deduplication and git does something similar with size and far weaker SHA-1. I would add a test for SHA-256 collisions, but no one has seemed to find a working example yet.

0 - https://github.com/ttkb-oss/dedup

[gblargg]

How much time is saved by not comparing full file contents? Given that this is a tool some people will only run occasionally, having it take 30 seconds instead of 15 is a small price to pay for ensuring it doesn't treat two differing files as equal.

[xanderlewis]

Same size, same first and last bytes, and same SHA-256.

…and you’re not worried about shark attacks, are you?

[gblargg]

I just don't understand why one would intentionally throw chance into a tool that one wants to be 100% robust. It's baffling.

[wink]

FWIW, when I wrote a tool like this I used same size + some hash function, not MD5 but maybe SHA1, don't remember. First and last bytes is a good idea, didn't think of that.

[pornel]

Reading just the first byte is probably wasting a read of the whole block.

Hashing the whole file after that is wasteful. You need to read (and hash) only as much as needed to demonstrate uniqueness of the file in the set.

The tree concept can be extended to every byte in the file:

https://github.com/kornelski/dupe-krill?tab=readme-ov-file#n...

[jonhohle]

Yeah, there is definitely some merit to more efficient hashing. Trees with a lot of duplicates require a lot of hashing, but hashing the entire file would be required regardless of whether partial hashes or done or not.

I have one data set where `dedup` was 40% faster than `dupe-krill` and another where `dupe-drill` was 45% faster than `dedup`.

`dupe-krill` uses blake3, which last I checked, was not hardware accelerated on M series processors. What's interesting is that because of hardware acceleration, `dedup` is mostly CPU-idle, waiting on the hash calculation, while `dupe-krill` is maxing out 3 cores.

Thanks for the link!

[unclebucknasty]

>which is probably really unique

Wonder what the distribution is here, on average? I know certain file types tend to cluster in specific ranges.

>maybe it’s better to just compare byte by byte? You’ll have to read the whole file to generate the hash

Definitely, for comparing any two files. But, if you're searching for duplicates across the entire disk, then you're theoretically checking each file multiple times, and each file is checked against multiple times. So, hashing them on first pass could conceivably be more efficient.

>if you just compare the bytes there is no chance of hash collision

You could then compare hashes and, only in the exceedingly rare case of a collision, do a byte-by-byte comparison to rule out false positives.

But, if your first optimization (the file size comparison) really does dramatically reduce the search space, then you'd also dramatically cut down on the number of re-comparisons, meaning you may be better off not hashing after all.

You could probably run the file size check, then based on how many comparisons you'll have to do for each matched set, decide whether hashing or byte-by-byte is optimal.

[pornel]

> exceedingly rare

To have a mere one in a billion chance of getting a SHA-256 collision, you'd need to spend 160 million times more energy than the total annual energy production on our planet (and that's assuming our best bitcoin mining efficiency, actual file hashing needs way more energy).

The probability of a collision is so astronomically small, that if your computer ever observed a SHA-256 collision, it would certainly be due to a CPU or RAM failure (bit flips are within range of probabilities that actually happen).

[unclebucknasty]

You know, I've been hearing people warn of handling potential collisions for years and knew the odds were negligible, but never really delved into it in any practical sense.

Context is everything.

[karparov]

This can be done much faster and safer.

You can group all files into buckets, and as soon as a bucket is empty, discard it. If in the end there are still files in the same bucket, they are duplicates.

Initially all files are in the same bucket.

You now iterate over differentiators which given two files tell you whether they are maybe equal or definitely not equal. They become more and more costly but also more and more exact. You run the differentiator on all files in a bucket to split the bucket into finer equivalence classes.

For example:

* Differentiator 1 is the file size. It's really cheap, you only look at metadata, not the file contents.

* Differentiator 2 can be a hash over the first file block. Slower since you need to open every file, but still blazingly fast and O(1) in file size.

* Differentiator 3 can be a hash over the whole file. O(N) in file size but so precise that if you use a cryptographic hash then you're very unlikely to have false positives still.

* Differentiator 4 can compare files bit for bit. Whether that is really needed depends on how much you trust collision resistance of your chosen hash function. Don't discard this though. Git got bitten by this.

[jonhohle]

Not surprisingly, differentiator 2 can just be the first byte (or machine word). Differentiator 3 can be the last byte (or word). At that point, 99.99% (in practice more 9s) of files are different and you’re read at most 2 blocks per file. I haven’t figured out a good differentiator 3 prior to hashing, but it’s already so rare, that it’s not worth it, in my experience.

[rzzzt]

I experimented with a similar, "hardlink farm"-style approach for deduplicated, browseable snapshots. It resulted in a small bash script which did the following:

- compute SHA256 hashes for each file on the source side

- copy files which are not already known to a "canonical copies" folder on the destination (this step uses the hash itself as the file name, which makes it easy to check if I had a copy from the same file earlier)

- mirror the source directory structure to the destination

- create hardlinks in the destination directory structure for each source file; these should use the original file name but point to the canonical copy.

Then I got too scared to actually use it :)

[kccqzy]

Hard links are not a suitable alternative here. When you deduplicate files, you typically want copy-on-write: if an app writes to one file, it should not change the other. Because of this, I would be extremely scared to use anything based on hard links.

In any case, a good design is to ask the kernel to do the dedupe step after user space has found duplicates. The kernel can double-check for you that they are really identical before doing the dedupe. This is available on Linux as the ioctl BTRFS_IOC_FILE_EXTENT_SAME.

[rzzzt]

It was for me. I was using rsync with "--link-dest" earlier for this purpose, but that only works if the file is present in consecutive backups. I wanted to have the option of seeing a potentially different subset of files for each backup and saving disk space at the same time.

Restic and Borg can do this at the block level, which is more effective but requires the tool to be installed when I want to check out something.

[620gelato]

Oh the sha-256 hashes are precisely what I used for a quick script I put together to parse through various backups of my work laptop in different places (tool changes and laziness). I had 10 different backups going back 4 years, and I wanted to make sure I - 1) preserved all unique files, 2) preserve the latest folder structure they showed up in.

Using sha256 was a no-brainer, at least for me.

[pmarreck]

xxHash (or xxh3 which I believe is even faster) is massively faster than SHA256 at the cost of security, which is unnecessary here.

Of course, engineering being what it is, it's possible that only one of these has hardware support and thus might end up actually being faster in realtime.

[PhilipRoman]

Blake3 is my favorite for this kind of thing. It's a cryptographic hash (maybe not the world's strongest, but considered secure), and also fast enough that in real world scenarios it performs just as well as non-crypto hashes like xx.

[f1shy]

I think the prob. is not so low. I remember reading here about a person getting a foto of another chat in a chat application, which was using sha in the background. I do not recall all the details, it is improbable, but possible.

[sgerenser]

LOL nope, I seriously doubt that was the result of a SHA256 collision.

[kittoes]

The probability is truly, obscenely, low. If you read about a collision then you surely weren't reading about SHA256.

https://crypto.stackexchange.com/questions/47809/why-havent-...

[amelius]

Or just use whatever algorithm rsync uses.

[diegs]

This reminds me of https://en.wikipedia.org/wiki/Venti_(software) which was a content-addressible filesystem which used hashes for de-duplication. Since the hashes were computed at write time, the performance penalty is amortized.

[w4yai]

I'd hash the first 1024 bytes of all files, and starts from there is any collision. That way you don't need to hash the whole (large) files, but only those with same hashes.

[amelius]

I suspect that bytes near the end are more likely to be different (even if there may be some padding). For example, imagine you have several versions of the same document.

Also, use the length of the file for a fast check.

[kstrauser]

At that point, why hash them instead of just using the first 1024 bytes as-is?

[borland]

In order to check if a file is a duplicate of another, you need to check it against _every other possible file_. You need some kind of "lookup key".

If we took the first 1024 bytes of each file as the lookup key, then our key size would be 1024 bytes. If you have 1 million files on your disk, then that's 128MB of ram just to store all the keys. That's not a big deal these days, but it's also annoying if you have a bunch of files that all start with the same 1024 bytes -- e.g. perhaps all the photoshop documents start with the same header. You'd need a 2-stage comparison, where you first match the key (1024 bytes) and then do a full comparison to see if it really matches.

Far more efficient - and less work - If you just use a SHA256 of the file's contents. That gets you a much smaller 32 byte key, and you don't need to bother with 2-stage comparisons.

[kstrauser]

I understand the concept. My main point is that it's probably not a huge advantage to store hashes of the first 1KB, which requires CPU to calculate, over just the raw bytes, which requires storage. There's a tradeoff either way.

I don't think it would be far more efficient to do hash the entire contents though. If you have a million files storing a terabyte of data, the 2 stage comparison would read at most 1GB (1 million * 1KB) of data, and less for smaller files. If you do a comparison of the whole hashed contents, you have to read the entire 1TB. There are a hundred confounding variables, for sure. I don't think you could confidently estimate which would be more efficient without a lot of experimenting.

[philsnow]

If you're going to keep partial hashes in memory, may as well align it on whatever boundary is the minimal block/sector size that your drives give back to you. Hashing (say) 8kB takes less time than it takes to fetch it from SSD (much less disk), so if you only used the first 1kB, you'd (eventually) need to re-fetch the same block to calculate the hash for the rest of the bytes in that block.

... okay, so as long as you always feed chunks of data into your hash in the same deterministic order, it doesn't matter for the sake of correctness what that order is or even if you process some bytes multiple times. You could hash the first 1kB, then the second-through-last disk blocks, then the entire first disk block again (double-hashing the first 1kB) and it would still tell you whether two files are identical.

If you're reading from an SSD and seek times don't matter, it's in fact probable that on average a lot of files are going to differ near the start and end (file formats with a header and/or footer) more than in the middle, so maybe a good strategy is to use the first 32k and the last 32k, and then if they're still identical, continue with the middle blocks.

In memory, per-file, you can keep something like

  - the length
  - h(block[0:4])
  - h(block[0:4] | block[-5:])
  - h(block[0:4] | block[-5:] | block[4:32])
  - h(block[0:4] | block[-5:] | block[4:128])
  - ...
  - h(block[0:4] | block[-5:] | block[4:])

etc, and only calculate the latter partial hashes when there is a collision between earlier ones. If you have 10M files and none of them have the same length, you don't need to hash anything. If you have 10M files and 9M of them are copies of each other except for a metadata tweak that resides in the last handful of bytes, you don't need to read the entirety of all 10M files, just a few blocks from each.

A further refinement would be to have per-file-format hashing strategies... but then hashes wouldn't be comparable between different formats, so if you had 1M pngs, 1M zips, and 1M png-but-also-zip quine files, it gets weird. Probably not worth it to go down this road.

[sedatk]

Probably because you need to keep a lot of those in memory.

[kstrauser]

I suspect that a computer with so many files that this would be useful probably has a lot of RAM in it, at least in the common case.

[sedatk]

But you need to constantly process them too, not just store them.

[smusamashah]

And why first 1024, can pick from predefined points.

[f1shy]

Depending on the medium, the penalty of reading single bytes in sparse locations could be comparable with reading the whole file. Maybe not a big win.

[williamsmj]

Deleted comment based on a misunderstanding.

[Sohcahtoa82]

> This tool simply identifies files that point at literally the same data on disk because they were duplicated in a copy-on-write setting.

You misunderstood the article, as it's basically doing the opposite of what you said.

This tool finds duplicate data that is specifically not duplicated via copy-on-write, and then turns it into a copy-on-write copy.

[williamsmj]

Fair. Deleted.