Ask Slashdot: How Do I De-Dupe a System With 4.2 Million Files?

First time accepted submitter jamiedolan writes "I've managed to consolidate most of my old data from the last decade onto drives attached to my main Windows 7 PC. Lots of files of all types from digital photos & scans to HD video files (also web site backup's mixed in which are the cause of such a high number of files). In more recent times I've organized files in a reasonable folder system and have an active / automated backup system. The problem is that I know that I have many old files that have been duplicated multiple times across my drives (many from doing quick backups of important data to an external drive that later got consolidate onto a single larger drive), chewing up space. I tried running a free de-dup program, but it ran for a week straight and was still 'processing' when I finally gave up on it. I have a fast system, i7 2.8Ghz with 16GB of ram, but currently have 4.9TB of data with a total of 4.2 million files. Manual sorting is out of the question due to the number of files and my old sloppy filing (folder) system. I do need to keep the data, nuking it is not a viable option.

Comments

CRC

[Spazmania]

Do a CRC32 of each file. Write to a file one per line in this order: CRC, directory, filename. Sort the file by CRC. Read the file linearly doing a full compare on any file with the same CRC (these will be adjacent in the file).

Re:CRC

[Anonymous Coward]

s/CRC32/sha1 or md5, you won't be CPU bound anyway.

Re:CRC

[Kral_Blbec]

Or just by file size first, then do a hash. No need to compute a hash to compare a 1mb file and a 1kb file.

Re:CRC

[igb]

That involves reading every byte. It would be faster to read the bytecount of each file, which doesn't involve reading the files themselves as that metadata is available, and then exclude from further examination all the files which have unique sizes. You could then read the first block of each large file, and discard all the files that have unique first blocks. After that, CRC32 (or MD5 or SHA1 --- you're going to be disk-bound anyway) and look for duplicates that way.

Re:CRC

[Zocalo]

No. No. No. Blindly CRCing every file is probably what took so long on the first pass and is a terribly inefficient way of de-duplicating files.

There is absolutely no point in generating CRCs of files unless they match on some other, simpler to compare characteristic like file size. The trick is to break the problem apart into smaller chunks. Start with the very large files, they exact size break to use it'll depend on the data set, but as the poster mentioned video file say everything over 1GB to start. Chances are you can fully de-dupe your very large files manually based on nothing more than a visual inspection of names and file sizes in little more time than it takes to find them all in the first place. You can then exclude those files from further checks, and more importantly, from CRC generation.

After that, try and break the problem down into smaller chunks. Whether you are sorting on size, name or CRC, it's quicker to do so when you only have a few hundred thousand files rather than several million. Maybe do another size constrained search; 512MB-1GB, say. Or if you have them, look for duplicated backups files in the form of ZIP files, or whatever archive format(s), you are using based on their extension - that also saves you having to expand and examine the contents of multiple archive files. Similarly, do a de-dupe of just the video files by extensions as these should again lend themselves to rapid manual sorting without having to generate CRCs for many GB of data. Another grouping to consider might be to at least try and get all of the website data, or as much of is as you can, into one place and de-dupe that, and consider whether you really need multiple archival copies of a site, or whether just the latest/final revision will do.

By the time you've done all that, including moving the stuff that you know is unique out of the way and into a better filing structure as you go, the remainder should be much more manageable for a single final pass. Scan the lot, identify duplicates based on something simple like the file size and, ideally, manually get your de-dupe tool to CRC only those groups of identically sized files that you can't easily tell apart like bunches of identically sized word processor or image files with cryptic file names.

Re:CRC

[caluml]

Exactly. What I do is this:

1. Compare filesizes.
2. When there are multiple files with the same size, start diffing them. I don't read the whole file to compute a checksum - that's inefficient with large files. I simply read the two files byte by byte, and compare - that way, I can quit checking as soon as I hit the first different byte.

Source is at https://github.com/caluml/finddups - it needs some tidying up, but it works pretty well.

git clone, and then mvn clean install.

Re:CRC

[Anonymous Coward]

If you get a linux image running (say in a livecd or VM) that can access the file system then fdupes is built to do this already. Various output format/recursion options.

From the man page:
DESCRIPTION
              Searches the given path for duplicate files. Such files are found by
              comparing file sizes and MD5 signatures, followed by a byte-by-byte
              comparison.

Re:CRC

[bzipitidoo]

Part 2 of your method will quickly bog down if you run into many files that are the same size. Takes (n choose 2) comparisons, for a problem that can be done in n time. If you have 100 files all of one size, you'll have to do 4950 comparisons. Much faster to compute and sort 100 checksums.

Also, you don't have to read the whole file to make use of checksums, CRCs, hashes and the like. Just check a few pieces likely to be different if the files are different, such as the first and last 2000 bytes. Then for those files with matching parts, check the full files.

Re:CRC

[blueg3]

b) Looking at the first few bytes of files with the same size.

Note that there's no reason to only look at the first few bytes. On spinning disks, any read smaller than about 16K will take the same amount of time. Comparing two 16K chunks takes zero time compared to how long it takes to read them from disk.

You could, for that matter, make it a 3-pass system that's pretty fast:
a) get all file sizes; remove all files that have unique sizes
b) compute the MD5 hash of the first 16K of each file; remove all files that have unique (size, header-hash) pairs
c) compute the MD5 hash of the whole file; remove all files that have unique (size, hash) pairs

Now you have a list of duplicates.

Don't forget to eliminate all files of zero length in step (a). They're trivially duplicates but shouldn't be deduplicated.

Re:CRC

[b4dc0d3r]

This was theorized by one of the RSA guys (Rivest, if I'm not mistaken). I helped support a system that identified files by CRC32, as a lot of tools did back then. As soon as we got to about 65k files (2^16), we had two files with the same CRC32.

Let me say, CRC32 is a very good algorithm. So good, I'll tell you how good. It is 4 bytes long, which means in theory you can change any 4 bytes of a file and get a CRC32 collision, unless the algorithm distributes them randomly, in which case you will get more or less.

I naively tried to reverse engineer a file from a known CRC32. Optimized and recursive, on a 333 mHz computer, it took 10 minutes to generate the first collision. Then every 10 minutes or so. Every 4 bytes (last 4, last 5 with the original last byte, last 6 with original last 2 bytes, etc) there was a collision.

Compare file sises first, not CRC32. The s^16 estimate is not only mathematically proven, but also in the big boy world. I tried to move the community towards another hash.

CRC32 *and* filesize are a great combination. File size is not included in the 2^16 estimate. I have yet to find two files in the real world, in the same domain (essentially type of file), with the same size and CRC32.

Be smart, use the right tool for the job. First compare file size (ignoring things like mp3 ID3 tags, or other headers). Then do two hashes of the contents - CRC32 and either MD5 or SHA1 (again ignoring well-known headers if possible). Then out of the results, you can do a byte for byte comparison, or let a human decide.

This is solely to dissuade CRC32 based identification. After all, it was designed for error detection, not identification. For a 4-byte file, my experience says CCITT standard CRC32 will work for identification. For 5 byte files, you can have two bytes swapped and possibly have the same result. The longer the file, the less likely it is to be unique.

Be smart, use size and two or more hashes to identify files. And even then, verify the contents. But don't compute hashes on every file - the operating system tells you file size as you traverse the directories, so start there.

Re:CRC

[BasilBrush]

Someone who's technical expertise is in areas other than writing script files. There are technical jobs other than being a sysop you know.

Re:CRC

[iluvcapra]

First compare file size (ignoring things like mp3 ID3 tags, or other headers).

I once had to write an audio file de-deuplicator; one of the big problems was you would ignore the metadata and the out-of-band data when you did the comparisons, but you always had to take this stuff into account when you were deciding which version of a file to keep -- you didn't want to delete two copies f a file with all the tags filled out and keep the one that was naked.

My de-duper worked like everyone here is saying -- it cracked open wav and aiff (and Sound Designer 2) files, captured their sample count and sample format into a sqlite db, did a couple of big joins and then did some SHA1 hashes of likely suspects. All of this worked great, but once I had the list I had the epiphany that the real problem of these tools is the resolution and how you make sure you're doing exactly what the user wants.

How do you decide which one to keep? You can just do hard links, but...

• The users I was working with were very uncomfortable with hard links, they didn't really understand the concept and were concerned that it made it difficult to know if you were "really" throwing something away when you dragged something to the trash. (It's stupid but it was their box.)
• Our existing backup/archival software wouldn't do the right thing with hard links, so it'd save no space on the tapes.
• Our audio workstation software wouldn't read audio off of files that were hard links on OS X (because hard links on OSX aren't really hard links, I believe our audio workstation vendor have since resolved this).

But let's say you can do hard links, no problem. How do you decide which instance of the file is to be kept, if you've only compared the "real" content of the file and ignored metadata? You could just give the user a big honking list of every set of files that are duplicates -- two here, three here, six here, and then let them go through and elect which one will be kept, but that's a mess and 99% of the time they're going to select a keeper on the basis of which part of the directory tree it's in. So, you need to do a rule system or a preferential ranking of parts of the directory hierarchy that tell the system "keep files you find here." Now, the files will also have metadata, so you also have to preferentially rank the files on the basis of its presence -- you might also rank files higher if your guy did the metadata tagging, because things like audio descriptions are often done with a specialized jargon that can be specific to a particular house.

Also, it'd be very common to delete a file from a directory containing an editor's personal library, and replacing it with a hard link to a file in the company's main library -- several people would have copies of the same commercial sound, or an editor would be the recordist of a sound that was subsequently sold to a commercial library, or whatever. Is it a good policy to replace his file with a hardlink to a different one, particularly if they differ in the metadata? Directories on a volume are often controlled by different people with different policies and proprietary interest to the files -- maybe the company "owns" everything, but it still can create a lot of internal disputes if files in a division or individual project's library folder starting getting their metadata changed, on account of being replaced with a hard link to a "better" file in the central repository. We can agree not to de-dup these, but it's more rules and exceptions that have to be made.

Once you have to list of duplicates, and maybe the rules, do you just go and delete, or do you give the user a big list to review? And, if upon review, he makes one change to one duplicate instance, it'd be nice to have that change intelligently reflected on the others. The rules have to be applied to the dupe list interactively and changes have to be reflected in the same way, otherwise it becomes a miserable experience for the user to de-dupe 1M files over 7 terabytes. The resolution of duplicates is the hard part, the finding of dupes is relatively easy.

Re:CRC

[xigxag]

With 4.2 million files, given the probability of SHA-1 collisions plus the birthday paradox and there will be around 500 SHA-1 collisions which are not duplicates.

That's totally, completely wrong. The birthday problem isn't a breakthrough concept, and the probability of random SHA-1 collisions is therefore calculated with it in mind. The number is known to be 1/2^80. This is straightforwardly derived from the total number of SHA-1 values, 2^160, which is then immensely reduced by the birthday paradox to 2^80 expected hashes required for a collision. This means that a hard drive with 2^80 or 1,208,925,819,614,629,174,706,176 files would have on average ONE collision. Note that this is a different number than the number of hashes one has to generate for a targeted cryptographic SHA-1 attack, which with best current theory is on the order of 2^51 for the full 80-round SHA-1, although as Goaway has pointed out, no such collision has yet been found.

Frankly I'm at a loss as to how you arrived at 500 SHA-1 collisions out of 4.2 million files. That's ludicrous. Any crypto hashing function with such a high collision rate would be useless. Much worse than MD5, even.

Re:CRC

[Surt]

$19.95 for a beta of something you can whip up in an hour of shell scripting.
If the poster were you, they wouldn't have had to 'ask slashdot'.

Re:ZFS

[smash]

To clarify - no this will not remove duplicate references to the data. The files ystem will remain in tact. However it will perform block level dedupe of the data which will recover your space. Duplicate references aren't necessarily a bad thing anyway, as if you have any sort of content index (memory, code, etc) that refers to data in a particular location, it will continue to work. However the space will be recovered.

There are tools for this

[Anonymous Coward]

If you don't mind booting Linux (a live version will do), fdupes has been fast enough for my needs and has various options to help you when multiple collisions occur. For finding similar images with non-identical checksums, findimagedupes will work, although it's obviously much slower than a straight 1-to-1 checksum comparison.

YMMV

Simple dedupe algorithm

[Anonymous Coward]

Delete all files but one. The remaining file is guaranteed unique!

Don't waste your time.

[Fuzzums]

if you really want, sort, order and index it all, but my suggestion would be different.

If you didn't need the files in the last 5 years, you'll probably never need them at all.
Maybe one or two. Make one volume called OldSh1t, index it, and forget about it again.

Really. Unless you have a very good reason to un-dupe everything, don't.

I have my share of old files and dupes. I know what you're talking about :)
Well, the sun is shining. If you need me, I'm outside.

Prioritize by file size

[jwales]

Since the objective is to recover disk space, the smallest couple of million files are unlikely to do very much for you at all. It's the big files that are the issue in most situations.

Compile a list of all your files, sorted by size. The ones that are the same size and the same name are probably the same file. If you're paranoid about duplicate file names and sizes (entirely plausible in some situations), then crc32 or byte-wise comparison can be done for reasonable or absolute certainty. Presumably at that point, to maintain integrity of any links to these files, you'll want to replace the files with hard links (not soft links!) so that you can later manually delete any of the "copies" without hurting all the other "copies". (There won't be separate copies, just hard links to one copy.)

If you give up after a week, or even a day, at least you will have made progress on the most important stuff.

Manual work will have to be done

[Qbertino]

Your problem isn't unduping files in your archives, your problem is getting an overview of your data archives. If you'd have it, you wouldn't have dupes in the first place.

This is a larger personal project, but you should take it on, since it will be a good lesson in data organisation. I've been there and done that.

You should get a rough overview of what you're looking at and where to expect large sets of dupes. Do this by manually parsing your archives in broad strokes. If you want to automate dupe-removal, do so by de-duping smaller chunks of your archive. You will need extra CPU and storage - maybe borrow a box or two from friends and set up a batch of scripts you can run from Linux live CDs with external HDDs attached.

Most likely you will have to do some scripting or programming, and you will have to devise a strategy not only of dupe removal, but of merging the remaining skeletons of dirtrees. That's actually the tough part. Removing dupes takes raw processing power and can be done in a few weeks and brute force and a solid storage bandwidth.

Organising the remaining stuff is where the real fun begins. ... You should start thinking about what you are willing to invest and how your backup, versioning and archiving strategy should look in the end, data/backup/archive retrival included. The latter might even determine how you go about doing your dirtree diffs - maybe you want to use a database for that for later use.

Anyway you put it, just setting up a box in the corner and having a piece of software churn away for a few days, weeks or months won't solve your problem in the end. If you plan well, it will get you started, but that's the most you can expect.

As I say: Been there, done that.
I still have unfinished business in my backup/archiving strategy and setup, but the setup now is 2 1TB external USB3 drives and manual arsync sessions every 10 weeks or so to copy from HDD-1 to HDD-2 to have dual backups/archives. It's quite simple now, but it was a long hard way to clean up the mess of the last 10 years. And I actually was quite conservative about keeping my boxed tidy. I'm still missing external storage in my setup, aka Cloud-Storage, the 2012 buzzword for that, but it will be much easyer for me to extend to that, now that I've cleaned up my shit halfway.

Good luck, get started now, work in iterations, and don't be silly and expect this project to be over in less than half a year.

My 2 cents.

Use DROID 6

[mattpalmer1086]

There is a digital preservation tool called DROID (Digital Record Object Identification) which scans all the files you ask it to, identifying their file type. It can also optionally generate an MD5 hash of each file it scans. It's available for download from sourceforge (BSD license, requires Java 6, update 10 or higher).

http://sourceforge.net/projects/droid/

It has a fairly nice GUI (for Java, anyway!), and a command line if you prefer scripting your scan. Once you have scanned all your files (with MD5 hash), export the results into a CSV file. If you like, you can first also define filters to exclude files you're not interested in (e.g. small files could be filtered out). Then import the CSV file into your data anlaysis app or database of your choice, and look for duplicate MD5 hashes. Alternetively, DROID actually stores its results in an Apache Derby database, so you could just connect directly to that rather than export to CSV, if you have a tool that an work with Derby.

One of the nice things about DROID when working over large datasets is you can save the progress at any time, and resume scanning later on. It was built to scan very large government datastores (multiple Tb). It has been tested over several million files (this can take a week or two to process, but as I say, you can pause at any time, save or restore, although only from the GUI, not the command line).

Disclaimer: I was responsible for the DROID 4, 5 and 6 projects while working at the UK National Archives. They are about to release an update to it (6.1 I think), but it's not available just yet.

Only if you have 100 unique files

[HiggsBison]

If you have 100 files all of one size, you'll have to do 4950 comparisons.

You only have to do 4950 comparisons if you have 100 unique files.

What I do is pop the first file from the list, to use as a standard, and compare all the files with it, block by block. If a block fails to match, I give up on that file matching the standard. The files that don't match generally don't go very far, and don't take much time. For the ones that match, I would have taken all that time if I was using a hash method anyway. As for reading the standard file multiple times: It goes fast because it's in cache.

The ones that match get taken from the list. Obviously I don't compare the one which match with each other. That would be stupid.

Then I go back to the list and rinse/repeat until there are less than 2 files.

I have done this many times with a set of 3 million files which take up about 600GB.