pyFileFixity

pyFileFixity provides a suite of open source, cross-platform, easy to use and easy to maintain (readable code) to protect and manage data for long term storage/archival, and also test the performance of any data protection algorithm.

The project is done in pure-Python to meet those criteria, although cythonized extensions are available for core routines to speed up encoding/decoding, but always with a pure python specification available so as to allow long term replication.

Here is an example of what pyFileFixity can do:

On the left, this is the original image.

At the center, the same image but with a few symbols corrupted (only 3 in header and 2 in the rest of the file, which equals to 5 bytes corrupted in total, over 19KB which is the total file size). Only a few corrupted bytes are enough to make the image looks like totally unrecoverable, and yet we are lucky, because the image could be unreadable at all if any of the "magic bytes" were to be corrupted!

At the right, the corrupted image was repaired using pff header command of pyFileFixity. This repaired only the image header (ie, the first part of the file), so only the first 3 corrupted bytes were repaired, not the 2 bytes in the rest of the file, but we can see the image looks indistinguishable from the untampered original! And the best thing is that it only costed the generation of a "ecc repair file" for the header, which size is only a constant 3.3KB per file, regardless of the protected file's size!

This works because most files will store the most important information to read them at their beginning, also called "file's header", so repairing this part will almost always ensure the possibility to read the file (even if the rest of the file is still corrupted, if the header is safe, you can read it). This works especially well for images, compressed files, formatted documents such as DOCX and ODT, etc.

Of course, you can also protect the whole file, not only the header, using pyFileFixity's pff whole command. You can also detect any corruption using pff hash.

Quickstart

Runs on Python 3 up to Python 3.12-dev. PyPy 3 is also supported.

• To install or update on Python 3:

pip install --upgrade pyfilefixity

• For Python 2.7, the latest working version was v3.0.2:

pip install --upgrade pyfilefixity==3.0.2 reedsolo==1.7.0 unireedsolomon==1.0.5

• Once installed, the suite of tools can be accessed from a centralized interface script called pff which provides several subcommands, to list them:

pff --help

You should see:

usage: pff [-h]
           {hash,rfigc,header,header_ecc,hecc,whole,structural_adaptive_ecc,saecc,protect,repair,recover,repair_ecc,recc,dup,replication_repair,restest,resilience_tester,filetamper,speedtest,ecc_speedtest}
           ...

positional arguments:
  {hash,rfigc,header,header_ecc,hecc,whole,structural_adaptive_ecc,saecc,protect,repair,recover,repair_ecc,recc,dup,replication_repair,restest,resilience_tester,filetamper,speedtest,ecc_speedtest}
    hash (rfigc)        Check files integrity fast by hash, size, modification date or by data structure integrity.
    header (header_ecc, hecc)
                        Protect/repair files headers with error correction codes
    whole (structural_adaptive_ecc, saecc, protect, repair)
                        Protect/repair whole files with error correction codes
    recover (repair_ecc, recc)
                        Utility to try to recover damaged ecc files using a failsafe mechanism, a sort of recovery
                        mode (note: this does NOT recover your files, only the ecc files, which may then be used to
                        recover your files!)
    dup (replication_repair)
                        Repair files from multiple copies of various storage mediums using a majority vote
    restest (resilience_tester)
                        Run tests to quantify robustness of a file protection scheme (can be used on any, not just
                        pyFileFixity)
    filetamper          Tamper files using various schemes
    speedtest (ecc_speedtest)
                        Run error correction encoding and decoding speedtests

options:
  -h, --help            show this help message and exit

• Every subcommands provide their own more detailed help instructions, eg for the hash submodule:

pff hash --help

• To generate a monitoring database (to later check very fast which files are corrupted, but cannot repair anything but filesystem metadata):

pff hash -i "your_folder" -d "dbhash.csv" -g -f -l "log.txt"

Note: this also works for a single file, just replace "your_folder" by "your_file.ext".

• To update this monitoring database (check for new files, but does not remove files that do not exist anymore - replace --append with --remove for the latter):

pff hash -i "your_folder -d "dbhash.csv" --update --append

• Later, to check which files were corrupted:

pff hash -i "your_folder" -d "dbhash.csv" -l log.txt -s -e errors.csv

• To use this monitoring database to recover filesystem metadata such as files names and directory layout by filescraping from files contents:

pff hash -i "your_folder" -d "dbhash.csv" -l "log.txt" -o "output_folder" --filescraping_recovery

• To protect files headers with a file called hecc.txt:

pff header -i "your_folder" -d "hecc.txt" -l "log.txt" -g -f --ecc_algo 3

• To repair files headers and store the repaired files in output_folder:

pff header -i "your_folder" -d "hecc.txt" -o "output_folder" -l "log.txt" -c -v --ecc_algo 3

• To protect whole files with a file called ecc.txt:

pff whole -i "your_folder" -d "ecc.txt" -l "log.txt" -g -f -v --ecc_algo 3

• To repair whole files:

pff whole -i "your_folder" -d "ecc.txt" -o "output_folder" -l "log.txt" -c -v --ecc_algo 3

Note that header and whole can also detect corrupted files and even which blocks inside a file, but they are much slower than hash.

• To try to recover a damaged ecc file ecc.txt using an index file ecc.txt.idx (index file is generated automatically with ecc.txt):

pff recovery -i "ecc.txt" --index "ecc.txt.idx" -o "ecc_repaired.txt" -l "log.txt" -v -f

• To try to recover a damaged ecc file ecc.txt without an index file (you can tweak the -t parameter from 0.0 to 1.0, 1.0 producing many false positives):

pff recovery -i "ecc.txt" -o "ecc_repaired.txt" -l "log.txt" -v -f -t 0.4

• To repair your files using multiple duplicated copies that you have stored on different mediums:

pff dup -i "path/to/dir1" "path/to/dir2" "path/to/dir3" -o "path/to/output" --report "rlog.csv" -f -v

• If you have previously generated a rfigc database, you can use it to enhance the replication repair:

pff dup -i "path/to/dir1" "path/to/dir2" "path/to/dir3" -o "path/to/output" -d "dbhash.csv" --report "rlog.csv" -f -v

• To run tests on your recovery tools, you can make a Makefile-like configuration file and use the Resiliency Tester submodule:

pff restest -i "your_folder" -o "test_folder" -c "resiliency_tester_config.txt" -m 3 -l "testlog.txt" -f

• Internally, pff restest uses pff filetamper to tamper files with various schemes, but you can also use pff filetamper directly.
• To run speedtests of encoding/decoding error correction codes on your machine:

pff speedtest

• In case the pff command does not work, it can be replaced with python -m pyFileFixity.pff .

The problem of long term storage

Why are data corrupted with time? One sole reason: entropy. Entropy refers to the universal tendency for systems to become less ordered over time. Data corruption is exactly that: a disorder in bits order. In other words: the Universe hates your data.

Long term storage is thus a very difficult topic: it's like fighting with death (in this case, the death of data). Indeed, because of entropy, data will eventually fade away because of various silent errors such as bit rot or cosmic rays. pyFileFixity aims to provide tools to detect any data corruption, but also fight data corruption by providing repairing tools.

The only solution is to use a principle of engineering that is long known and which makes bridges and planes safe: add some redundancy.

There are only 2 ways to add redundancy:

• the simple way is to duplicate the object (also called replication), but for data storage, this eats up a lot of storage and is not optimal. However, if storage is cheap, then this is a good solution, as it is much faster than encoding with error correction codes. For replication to work, at least 3 duplicates are necessary at all times, so that if one fails, it must replaced asap. As sailors say: "Either bring 1 compass or 3 compasses, but never two, because then you won't know which one is correct if one fails." Indeed, with 3 duplicates, if you frequently monitor their integrity (eg, with hashes), then if one fails, simply do a majority vote: the bit value given by 2 of the duplicates is probably correct.
• the second way, the optimal tools ever invented to recover from data corruption, are the error correction codes (forward error correction), which are a way to smartly produce redundant codes from your data so that you can later repair your data using these additional pieces of information (ie, an ECC generates n blocks for a file cut in k blocks (with k < n), and then the ecc code can rebuild the whole file with (at least) any k blocks among the total n blocks available). In other words, you can correct up to (n-k) erasures. But error correcting codes can also detect and repair automatically where the errors are (fully automatic data repair for you !), but at the cost that you can then only correct (n-k)/2 errors.

Error correction can seem a bit magical, but for a reasonable intuition, it can be seen as a way to average the corruption error rate: on average, a bit will still have the same chance to be corrupted, but since you have more bits to represent the same data, you lower the overall chance to lose this bit.

The problem is that most theoretical and pratical works on error correcting codes has been done almost exclusively on channel transmission (such as 4G, internet, etc.), but not on data storage, which is very different for one reason: whereas in a channel we are in a spatial scheme (both the sender and the receiver are different entities in space but working at the same timescale), in data storage this is a temporal scheme: the sender was you storing the data on your medium at time t, and the receiver is again you but now retrieving the data at time t+x. Thus, the sender does not exist anymore, thus you cannot ask the sender to send again some data if it's too much corrupted: in data storage, if a data is corrupted, it's lost for good, whereas in channel theory, parts of the data can be submitted again if necessary.

Some attempts were made to translate channel theory and error correcting codes theory to data storage, the first being Reed-Solomon which spawned the RAID schema. Then CIRC (Cross-interleaved Reed-Solomon coding) was devised for use on optical discs to recover from scratches, which was necessary for the technology to be usable for consumers. Since then, new less-optimal but a lot faster algorithms such as LDPC, turbo-codes and fountain codes such as RaptorQ were invented (or rediscovered), but they are still marginally researched for data storage.

This project aims to, first, implement easy tools to evaluate strategies (filetamper.py) and file fixity (ie, detect if there are corruptions), and then the goal is to provide an open and easy framework to use different kinds of error correction codes to protect and repair files.

Also, the ecc file specification is made to be simple and resilient to corruption, so that you can process it by your own means if you want to, without having to study for hours how the code works (contrary to PAR2 format).

In practice, both approaches are not exclusive, and the best is to combine them: protect the most precious data with error correction codes, then duplicate them as well as less sensitive data across multiple storage mediums. Hence, this suite of data protection tools, just like any other such suite, is not sufficient to guarantee your data is protected, you must have an active (but infrequent and hence not time consuming) data curation strategy that includes regularly checking your data and replacing copies that are damaged every few years.

For a primer on storage mediums and data protection strategies, see this post I wrote.

Why not just use RAID ?

RAID is clearly insufficient for long-term data storage, and in fact it was primarily meant as a cheap way to get more storage (RAID0) or more availability (RAID1) of data, not for archiving data, even on a medium timescale:

• RAID 0 is just using multiple disks just like a single one, to extend the available storage. Let's skip this one.
• RAID 1 is mirroring one disk with a bit-by-bit copy of another disk. That's completely useless for long term storage: if either disk fails, or if both disks are partially corrupted, you can't know what are the correct data and which aren't. As an old saying goes: "Never take 2 compasses: either take 3 or 1, because if both compasses show different directions, you will never know which one is correct, nor if both are wrong." That's the principle of Triplication.
• RAID 5 is based on the triplication idea: you have n disks (but least 3), and if one fails you can recover n-1 disks (resilient to only 1 disk failure, not more).
• RAID 6 is an extension of RAID 5 which is closer to error-correction since you can correct n-k disks. However, most (all?) currently commercially available RAID6 devices only implements recovery for at most n-2 (2 disks failures).
• In any case, RAID cannot detect silent errors automatically, thus you either have to regularly scan, or you risk to lose some of your data permanently, and it's far more common than you can expect (eg, with RAID5, it is enough to have 2 silent errors on two disks on the same bit for the bit to be unrecoverable). That's why a limit of only 1 or 2 disks failures is just not enough.
• Finally, it's worth noting that hard drives do implement ECC codes to be resilient against bad sectors (otherwise we would lose data all the time!), but they only have limited corrective capacity, mainly because the ECC code is short and not configurable.

On the opposite, ECC can correct n-k disks (or files). You can configure n and k however you want, so that for example you can set k = n/2, which means that you can recover all your files from only half of them! (once they are encoded with an ecc file of course).

There also are new generation RAID solutions, mainly software based, such as SnapRAID or ZFS, which allow you to configure a virtual RAID with the value n-k that you want. This is just like an ecc file (but a bit less flexible, since it's not a file but a disk mapping, so that you can't just copy it around or upload it to a cloud backup hosting). In addition to recover (n-k) disks, they can also be configured to recover from partial, sectors failures inside the disk and not just the whole disk (for a more detailed explanation, see Plank, James S., Mario Blaum, and James L. Hafner. "SD codes: erasure codes designed for how storage systems really fail." FAST. 2013.).

The other reason RAID is not adapted to long-term storage, is that it supposes you store your data on hard-drives exclusively. Hard drives aren't a good storage medium for the long term, for two reasons:

In the end, it's a lot better to just separate the storage medium of data, with the reading instrument.

We will talk later about what storage mediums can be used instead.

Applications included

The pyFileFixity suite currently include the following pure-python applications:

• rfigc.py (subcommand: hash), a hash auditing tool, similar to md5deep/hashdeep, to compute a database of your files along with their metadata, so that later you can check if they were changed/corrupted.
• header_ecc.py (subcommand: header), an error correction code using Reed-Solomon generator/corrector for files headers. The idea is to supplement other more common redundancy tools such as PAR2 (which is quite reliable), by adding more resiliency only on the critical parts of the files: their headers. Using this script, you can significantly higher the chance of recovering headers, which will allow you to at least open the files.
• structural_adaptive_ecc.py (subcommand: whole), a variable error correction rate encoder (kind of a generalization of header_ecc.py). This script allows to generate an ecc file for the whole content of your files, not just the header part, using a variable resilience rate: the header part will be the most protected, then the rest of each file will be progressively encoded with a smaller and smaller resilience rate. The assumption is that important information is stored first, and then data becomes less and less informative (and thus important, because the end of the file describes less important details). This assumption is very true for all compressed kinds of formats, such as JPG, ZIP, Word, ODT, etc...
• repair_ecc.py (subcommand: recovery), a script to repair the structure (ie, the entry and fields markers/separators) of an ecc file generated by header_ecc.py or structural_adaptive_ecc.py. The goal is to enhance the resilience of ecc files against corruption by ensuring that their structures can be repaired (up to a certain point which is very high if you use an index backup file, which is a companion file that is generated along an ecc file).
• filetamper.py (subcommand: filetamper) is a quickly made file corrupter, it will erase or change characters in the specified file. This is useful for testing your various protecting strategies and file formats (eg: is PAR2 really resilient against corruption? Are zip archives still partially extractable after corruption or are rar archives better? etc.). Do not underestimate the usefulness of this tool, as you should always check the resiliency of your file formats and of your file protection strategies before relying on them.
• replication_repair.py (subcommand: dup) takes advantage of your multiple copies (replications) of your data over several storage mediums to recover your data in case it gets corrupted. The goal is to take advantage of the storage of your archived files into multiple locations: you will necessarily make replications, so why not use them for repair? Indeed, it's good practice to keep several identical copies of your data on several storage mediums, but in case a corruption happens, usually you will just drop the corrupted copies and keep the intacts ones. However, if all copies are partially corrupted, you're stuck. This script aims to take advantage of these multiple copies to recover your data, without generating a prior ecc file. It works simply by reading through all your different copies of your data, and it casts a majority vote over each byte: the one that is the most often occuring will be kept. In engineering, this is a very common strategy used for very reliable systems such as space rockets, and is called "triple-modular redundancy", because you need at least 3 copies of your data for the majority vote to work (but the more the better).
• resiliency_tester.py (subcommand: restest) allows you to test the robustness of the corruption correction of the scripts provided here (or any other command-line app). You just have to copy the files you want to test inside a folder, and then the script will copy the files into a test tree, then it will automatically corrupt the files randomly (you can change the parameters like block burst and others), then it will run the file repair command-lines you supply and finally some stats about the repairing power will be generated. This allows you to easily and objectively compare different set of parameters, or even different file repair solutions, on the very data that matters to you, so that you can pick the best option for you.
• ecc_speedtest.py (subcommand: speedtest) is a simple error correction codes encoder/decoder speedtest. It allows to easily change parameters for the test. This allows to assess how fast your machine can encode/decode with the selected parameters, which can be especially useful to plan ahead for how many files you can reasonably plan to protect with error correction codes (which are time consuming).
• DEPRECATED: easy_profiler.py is just a quick and simple profiling tool to get you started quickly on what should be optimized to get more speed, if you want to contribute to the project feel free to propose a pull request! (Cython and other optimizations are welcome as long as they are cross-platform and that an alternative pure-python implementation is also available).

Note that all tools are primarily made for command-line usage (type pff <subcommand> --help to get extended info about the accepted arguments)

IMPORTANT: it is CRITICAL that you use the same parameters for correcting mode as when you generated the database/ecc files (this is true for all scripts in this bundle). Of course, some options must be changed: -g must become -c to correct, and --update is a particular case. This works this way on purpose for mainly two reasons: first because it is very hard to autodetect the parameters from a database file alone and it would produce lots of false positives, and secondly (the primary reason) is that storing parameters inside the database file is highly unresilient against corruption (if this part of the database is tampered, the whole becomes unreadable, while if they are stored outside or in your own memory, the database file is always accessible). Thus, it is advised to write down the parameters you used to generate your database directly on the storage media you will store your database file on (eg: if it's an optical disk, write the parameters on the cover or directly on the disk using a marker), or better memorize them by heart. If you forget them, don't panic, the parameters are always stored as comments in the header of the generated ecc files, but you should try to store them outside of the ecc files anyway.

For users: what are the advantages of pyFileFixity?

Pros:

• Open application and open specifications under the MIT license (you can do whatever you want with it and tailor it to your needs if you want to, or add better decoding procedures in the future as science progress so that you can better recover your data from your already generated ecc file).
• Highly reliable file fixity watcher: rfigc.py will tell you without any ambiguity using several attributes if your files have been corrupted or not, and can even check for images if the header is valid (ie: if the file can still be opened).
• Readable ecc file format (compared to PAR2 and most other similar specifications).
• Highly resilient ecc file format against corruption (not only are your data protected by ecc, the ecc file is protected too against critical spots, both because there is no header so that each track is independent and if one track is corrupted beyond repair then other ecc tracks can still be read, and a .idx file will be generated to repair the structure of the ecc file to recover all tracks).
• Very safe and conservative approach: the recovery process checks that the recovery was successful before committing a repaired block.
• Partial recovery allowed (even if a file cannot be completely recovered, the parts that can will be repaired and then the rest that can't be repaired will be recopied from the corrupted version).
• Support directory processing: you can encode an ecc file for a whole directory of files (with any number of sub-directories and depth).
• No limit on the number of files, and it can recursively protect files in a directory tree.
• Variable resiliency rate and header-only resilience, ensuring that you can always open your files even if partially corrupted (the structure of your files will be saved, so that you can use other softwares to repair beyond if this set of script is not sufficient to totally repair).
• Support for erasures (null bytes) and even errors-and-erasures, which literally doubles the repair capabilities. To my knowledge, this is the only freely available parity software that supports erasures.
• Display the predicted total ecc file size given your parameters, and the total time it will take to encode/decode.
• Your original files are still accessible as they are, protection files such as ecc files live alongside your original data. Contrary to other data protection schemes such as PAR2 which encode the whole data in par archive files that replace your original files and are not readable without decoding.
• Opensourced under the very permissive MIT licence, do whatever you want!

Cons:

• Cannot protect meta-data, such as folders paths. The paths are stored, but cannot be recovered (yet? feel free to contribute if you know how). Only files are protected. Thus if your OS or your storage medium crashes and truncate a whole directory tree, the directory tree can't be repaired using the ecc file, and thus you can't access the files neither. However, you can use file scraping to extract the files even if the directory tree is lost, and then use RFIGC.py to reorganize your files correctly. There are alternatives, see the chapters below: you can either package all your files in a single archive using DAR or ZIP (thus the ecc will also protect meta-data), or see DVDisaster as an alternative solution, which is an ecc generator with support for directory trees meta-data (but only on optical disks).
• Can only repair errors and erasures (characters that are replaced by another character), not deletion nor insertion of characters. However this should not happen with any storage medium (truncation can occur if the file bounds is misdetected, in this case pyFileFixity can partially repair the known parts of the file, but cannot recover the rest past the truncation, except if you used a resiliency rate of at least 0.5, in which case any message block can be recreated with only using the ecc file).
• Cannot recreate a missing file from other available files (except you have set a resilience_rate at least 0.5), contrary to Parchives (PAR1/PAR2). Thus, you can only repair a file if you still have it (and its ecc file!) on your filesystem. If it's missing, pyFileFixity cannot do anything (yet, this will be implemented in the future).

Note that the tools were meant for data archival (protect files that you won't modify anymore), not for system's files watching nor to protect all the files on your computer. To do this, you can use a filesystem that directly integrate error correction code capacity, such as ZFS.

Recursive/Relative Files Integrity Generator and Checker in Python (aka RFIGC)

Recursively generate or check the integrity of files by MD5 and SHA1 hashes, size, modification date or by data structure integrity (only for images).

This script is originally meant to be used for data archival, by allowing an easy way to check for silent file corruption. Thus, this script uses relative paths so that you can easily compute and check the same redundant data copied on different mediums (hard drives, optical discs, etc.). This script is not meant for system files corruption notification, but is more meant to be used from times-to-times to check up on your data archives integrity (if you need this kind of application, see avpreserve's fixity).

Example usage

• To generate the database (only needed once):

pff hash -i "your_folder" -d "dbhash.csv" -g

• To check:

pff hash -i "your_folder" -d "dbhash.csv" -l log.txt -s

• To update your database by appending new files:

pff hash -i "your_folder" -d "dbhash.csv" -u -a

• To update your database by appending new files AND removing inexistent files:

pff hash -i "your_folder" -d "dbhash.csv" -u -a -r

Note that by default, the script is by default in check mode, to avoid wrong manipulations. It will also alert you if you generate over an already existing database file.

Arguments

  -h, --help            show a help message and exit
  -i /path/to/root/folder, --input /path/to/root/folder
                        Path to the root folder from where the scanning will occ
ur.
  -d /some/folder/databasefile.csv, --database /some/folder/databasefile.csv
                        Path to the csv file containing the hash informations.
  -l /some/folder/filename.log, --log /some/folder/filename.log
                        Path to the log file. (Output will be piped to both the
stdout and the log file)
  -s, --structure_check
                        Check images structures for corruption?
  -e /some/folder/errorsfile.csv, --errors_file /some/folder/errorsfile.csv
                        Path to the error file, where errors at checking will be
 stored in CSV for further processing by other softwares (such as file repair so
ftwares).
  -m, --disable_modification_date_checking
                        Disable modification date checking.
  --skip_missing        Skip missing files when checking (useful if you split yo
ur files into several mediums, for example on optical discs with limited capacit
y).
  -g, --generate        Generate the database? (omit this parameter to check ins
tead of generating).
  -f, --force           Force overwriting the database file even if it already e
xists (if --generate).
  -u, --update          Update database (you must also specify --append or --rem
ove).
  -a, --append          Append new files (if --update).
  -r, --remove          Remove missing files (if --update).

  --filescraping_recovery          Given a folder of unorganized files, compare to the database and restore the filename and directory structure into the output folder.
  -o, --output          Path to the output folder where to output the files reorganized after --recover_from_filescraping.

Header Error Correction Code script

This script was made to be used in combination with other more common file redundancy generators (such as PAR2, I advise MultiPar). This is an additional layer of protection for your files: by using a higher resiliency rate on the headers of your files, you ensure that you will be probably able to open them in the future, avoiding the "critical spots", also called "fracture-critical" in redundancy engineering (where if you modify just one bit, your whole file may become unreadable, usually bits residing in the headers - in other words, a single blow makes the whole thing collapse, just like non-redundant bridges).

An interesting benefit of this approach is that it has a low storage (and computational) overhead that scales linearly to the number of files, whatever their size is: for example, if we have a set of 40k files for a total size of 60 GB, with a resiliency_rate of 30% and header_size of 1KB (we limit to the first 1K bytes/characters = our file header), then, without counting the hash per block and other meta-data, the final ECC file will be about 2 * resiliency_rate * number_of_files * header_size = 24.5 MB. This size can be lower if there are many files smaller than 1KB. This is a pretty low storage overhead to backup the headers of such a big number of files.

The script is pure-python as are its dependencies: it is thus completely cross-platform and open source. The default ecc algo (ecc_algo=3 uses reedsolo) also provides a speed-optimized C-compiled implementation (creedsolo) that will be used if available for the user's platform, so pyFileFixity should be fast by default. Alternatively, it's possible to use a JIT compiler such as PyPy, although this means that creedsolo will not be useable, so PyPy may accelerate other functions but slower ecc encoding/decoding.

Structural Adaptive Error Correction Encoder

This script implements a variable error correction rate encoder: each file is ecc encoded using a variable resiliency rate -- using a high constant resiliency rate for the header part (resiliency rate stage 1, high), then a variable resiliency rate is applied to the rest of the file's content, with a higher rate near the beginning of the file (resiliency rate stage 2, medium) which progressively decreases until the end of file (resiliency rate stage 3, the lowest).

The idea is that the critical parts of files usually are placed at the top, and data becomes less and less critical along the file. What is meant by critical is both the critical spots (eg: if you tamper only one character of a file's header you have good chances of losing your entire file, ie, you cannot even open it) and critically encoded information (eg: archive formats usually encode compressed symbols as they go along the file, which means that the first occurrence is encoded, and then the archive simply writes a reference to the symbol. Thus, the first occurrence is encoded at the top, and subsequent encoding of this same data pattern will just be one symbol, and thus it matters less as long as the original symbol is correctly encoded and its information preserved, we can always try to restore the reference symbols later). Moreover, really redundant data will be placed at the top because they can be reused a lot, while data that cannot be too much compressed will be placed later, and thus, corruption of this less compressed data is a lot less critical because only a few characters will be changed in the uncompressed file (since the data is less compressed, a character change on the not-so-much compressed data won't have very significant impact on the uncompressed data).

This variable error correction rate should allow to protect more the critical parts of a file (the header and the beginning of a file, for example in compressed file formats such as zip or jpg this is where the most importantly strings are encoded) for the same amount of storage as a standard constant error correction rate.

Of course, you can set the resiliency rate for each stage to the values you want, so that you can even do the opposite: setting a higher resiliency rate for stage 3 than stage 2 will produce an ecc that is greater towards the end of the contents of your files.

Furthermore, the currently designed format of the ecc file would allow two things that are not available in all current file ecc generators such as PAR2:

1. it allows to partially repair a file, even if not all the blocks can be corrected (in PAR2, a file is repaired only if all blocks can be repaired, which is a shame because there are still other blocks that could be repaired and thus produce a less corrupted file) ;

2. the ecc file format is quite simple and readable, easy to process by any script, which would allow other softwares to also work on it (and it was also done in this way to be more resilient against error corruptions, so that even if an entry is corrupted, other entries are independent and can maybe be used, thus the ecc is very error tolerant. This idea was implemented in repair_ecc.py but it could be extended, especially if you know the pattern of the corruption).

The script structural-adaptive-ecc.py implements this idea, which can be seen as an extension of header-ecc.py (and in fact the idea was the other way around: structural-adaptive-ecc.py was conceived first but was too complicated, then header-ecc.py was implemented as a working lessened implementation only for headers, and then structural-adaptive-ecc.py was finished using header-ecc.py code progress). It works, it was a quite well tested for my own needs on datasets of hundred of GB, but it's not foolproof so make sure you test the script by yourself to see if it's robust enough for your needs (any feedback about this would be greatly appreciated!).

ECC Algorithms

You can specify different ecc algorithms using the --ecc_algo switch.

For the moment, only Reed-Solomon is implemented, but it's universal so you can modify its parameters in lib/eccman.py.

Two Reed-Solomon codecs are available, they are functionally equivalent and thoroughly unit tested.

• --ecc_algo 1: use the first Reed-Solomon codec in galois field 2^8 of root 3 with fcr=1. This is the slowest implementation (but also the most easy code to understand).
• --ecc_algo 2: same as algo 1 but with a faster functions.
• --ecc_algo 3: use the second codec, which is the fastest. The generated ECC will be compatible with algo 1 and 2.
• --ecc_algo 4: also use the second, fastest RS codec, but with different parameters (US FAA ADSB UAT RS FEC norm), thus the generated ECC won't be compatible with algo 1 to 3. But do not be scared, the ECC will work just the same.

Note about speed: Also, use a smaller --max_block_size to greatly speedup the operations! That's the trick used to compute very quickly RS ECC on optical discs. You give up a bit of resiliency of course (because blocks are smaller, thus you protect a smaller number of characters per ECC. In the end, this should not change much about real resiliency, but in case you get a big bit error burst on a contiguous block, you may lose a whole block at once. That's why using RS255 is better, but it's very time consuming. However, the resiliency ratios still hold, so for any other case of bit-flipping with average-sized bursts, this should not be a problem as long as the size of the bursts is smaller than an ecc block.)

In case of a catastrophic event

TODO: write more here

In case of a catastrophic event of your data due to the failure of your storage media (eg: your hard drive crashed), then follow the following steps:

1- use dd_rescue to make a full bit-per-bit verbatim copy of your drive before it dies. The nice thing with dd_rescue is that the copy is exact, and also that it can retries or skip in case of bad sectors (it won't crash on your suddenly at half the process).

2- Use testdisk to restore partition or to copy files based on partition filesystem informations.

3- If you could not recover your files, you can try file scraping using photorec or plaso other similar tools as a last resort to extract data based only from files content (no filename, often uncorrect filetype, file boundaries may be wrong so some data may be cut off, etc.).

4- If you used pyFileFixity before the failure of your storage media, you can then use your pre-computed databases to check that files are intact (rfigc.py) and if they aren't, you can recover them (using header_ecc.py and structural_adaptive_ecc.py). It can also help if you recovered your files via data scraping, because your files will be totally unorganized, but you can use a previously generated database file to recover the full names and directory tree structure using rfigc.py --filescraping_recover.

Also, you can try to fix some of your files using specialized repairing tools (but remember that such tool cannot guarantee you the same recovering capacity as an error correction code - and in addition, error correction code can tell you when it has recovered successfully). For example:

• for tar files, you can use fixtar. Similar tools (but older): tarfix and tar-repair.
• for RAID mounting and recovery, you can use "Raid faster - recover better" (rfrb) tool by Sabine Seufert and Christian Zoubek: https://github.com/lrq3000/rfrb
• if your unicode strings were mangled (ie, you see weird symbols), try this script that will automatically demangle them: https://github.com/LuminosoInsight/python-ftfy
• to repair tabular (2D) data such as .csv, try Carpenter.
• tool to identify corrupted files in ddrescue images: ddrescue-ffile

Protecting directory tree meta-data

One main current limitation of pyFileFixity is that it cannot protect the directory tree meta-data. This means that in the worst case, if a silent error happens on the inode pointing to the root directory that you protected with an ecc, the whole directory will vanish, and all the files inside too. In less worst cases, sub-directories can vanish, but it's still pretty bad, and since the ecc file doesn't store any information about inodes, you can't recover the full path.

The inability to store these meta-data is because of two choices in the design:

1. portability: we want the ecc file to work even if we move the root directory to another place or another storage medium (and of course, the inode would change),
2. cross-platform compatibility: there's no way to get and store directory meta-data for all platforms, but of course we could implement specific instructions for each main platform, so this point is not really a problem.

To workaround this issue (directory meta-data are critical spots), other softwares use a one-time storage medium (ie, writing your data along with generating and writing the ecc). This way, they can access at the bit level the inode info, and they are guaranted that the inodes won't ever change. This is the approach taken by DVDisaster: by using optical mediums, it can compute inodes that will be permanent, and thus also encode that info in the ecc file. Another approach is to create a virtual filesystem specifically to store just your files, so that you manage the inode yourself, and you can then copy the whole filesystem around (which is really just a file, just like a zip file - which can also be considered as a mini virtual file system in fact) like rsbep.

Here the portability principle of pyFileFixity prevents this approach. But you can mimic this workaround on your hard drive for pyFileFixity to work: you just need to package all your files into one file. This way, you sort of create a virtual file system: inside the archive, files and directories have meta-data just like in a filesystem, but from the outside it's just one file, composed of bytes that we can just encode to generate an ecc file - in other words, we removed the inodes portability problem, since this meta-data is stored relatively inside the archive, the archive manage it, and we can just encode this info like any other stream of data! The usual way to make an archive from several files is to use TAR, but this will generate a solid archive which will prevent partial recovery. An alternative is to use DAR, which is a non-solid archive version of TAR, with lots of other features too. If you also want to compress, you can just use ZIP (with DEFLATE algorithm) your files (this also generates a non-solid archive). You can then use pyFileFixity to generate an ecc file on your DAR or ZIP archive, which will then protect both your files just like before and the directories meta-data too now.

Which storage medium to use

Since hard drives have a relatively short timespan (5-10 years, often less) and require regular plugging to an electrical outlet to keep the magnetic plates from decaying, other solutions are more advisable.

The medium I used to advise was optical disks (whether it's BluRay, DVD - not CDs!), because the reading instrument is distinct from the storage medium, and the technology (laser reflecting on bumps and/or pits) is kind of universal, so that even if the technology is lost one day (deprecated by newer technologies, so that you can't find the reading instrument anymore because it's not sold anymore), you can probably emulate a laser using some software to read your optical disk, just like what the CAMiLEON project did to recover data from the LaserDiscs of the BBC Domesday Project (see Wikipedia). BluRays have an estimated lifespan of 20-50 years depending on if they are "gold archival grade", whereas DVD should live up from 10-30 years. CDs are only required to live a minimum of 1 year up to 10 years max, hence are not fit for archival. Archival optimized optical discs such as M-Discs boast about being able to live up to 100 years, but there is no independent scientific backing of these claims currently. For more details, you can read a longer explanation I wrote with references on StackOverflow.

However, limitations of optical discs include their limited storage space, low transfer speed, and limited rewriteability.

A more convenient solution is to use magnetic tape, especially with an open standard such as Linear Tape Open (LTO), which ensures interoperability between manufacturers and hence also reduces cost because of competition. LTO works as a two components system: the tape drive, and the cartridges (with the magnetic bands). There are lots of versions of LTO, each generation improving on the previous one. LTO cartridges have a shorter lifespan than optical discs, being 15-30 years on average, but they are much more convenient to use:

• they provide extremely big storage space (one cartridge being several TB as of LTO-4, and the storage capacity approximately doubles every few years with every new version!),
• are fast to write (about 5h to write the full cartridge, speed increases with new versions so the total time to fill a cartridge stays about the same),
• the storage medium (cartridges) is also distinct from the reading/writing instrument (LTO tape drive),
• are easily rewriteable, although it is necessary to reformat to free up space, but the idea is that "full mirror backups" can be made regularly by overwriting an old tape.
• being an open standard, drives to read older versions 25 years old (LTO-1 is from 2000) are still available.
• 15-30 years of lifespan is still great for archival! But requires active curation (ie, checking cartridges every 5 years and making a full new copy on a new cartridge each decade should be largely sufficient).
• Cartridges are cheap: LTO7 cartridges allowing storage of up to 15 TB cost only 60 bucks brand new, often much less in refurbished (already used, but can be overwritten and reused). This is MUCH less expensive than hard drives.
• Fit for cold storage: unlike hard drives (using magnetic platters) and like optical discs, the cartridges do not need to be plugged to an electrical outlet regularly, the magnetic band does not decay without electrical current, so the cartridges can be cold stored in air-tight, temperature-proofed and humidity-proof containers, which can be stored off-site (fire-proof data recovery plan).
• Recovery of failed LTO cartridges is inexpensive and readily available, whereas recovering the magnetic signal from failed hard drives costs thousands of euros/dollars. LTO tapes are also fully compatible with DAR archives, improving chances of recovery with error correction codes and non-solid archives that can be partially recovered.

Sounds perfect, right? Well, nothing is, LTO also has several disadvantages:

• Initial cost of starting is very expensive: a brand new LTO drive of latest generations cost several thousand euros/dollars. Second-hand or refurbished drives of older generations are much less expensive, but they are difficult to setup, as it is unlikely you will find them in an all-in-one package, you will have to get the tape drive separately from the computer system to plug it to (more on that in the next section).
• Limited retrocompatibility: the LTO standard specifies that each generation of drives only need to support the current gen and one past gen. However, this is counterbalanced by the fact that the LTO standard is open, so anybody can make LTO drives, including in the future, and it is possible someday a manufacturer will make a LTO drive that supports multiple past generations (just like there are old tapes digitizers that can be connected in USB, for archival purposes). Until then, in practice, it means that ideally when upgrading your LTO system, you need to upgrade by one generation at a time, or if you get a drive of 2+ later gens, you need to keep or buy a drive of the older gen you had to read your tapes to then transfer to the latest gen you have. As of 2023, there are still LTO1 tape drives available for cheap in second-hand, a technology that was published in 2000 and already deprecated in 2001 by LTO2, so this shows that LTO tape drives of older generations should still be plentily available.
• LTO is a sequential technology: it is very fast to write and read sequentially, but if you want to download a specific file, the tape has to be fully read up to where the file is stored, contrary to hard drives with random access that can access in linear or sublinear time.
• (Old fixed issue) Before LTO-5, which introduced the LTFS standardized filesystem that allows mounting on any operating file system such as Windows, Linux and MacOS, the various LTO drives manufacturers used their own closed-source filesystems that were often incompatible with each others. Hence, make sure to get an LTO-5 drive or above to ensure future access to your long term archives.

Given all the above characteristics, LTO>=5 appears to be the best practical solution for long term archival, if coupled with an active (but infrequent) curation process.

There is however one exception: if you need to cold store the medium in a non temperate environment (outside of 10-40°C), then using optical discs may be more resilient, although LTO cartridges should also be able to sustain a wider range of temperature but you need to wait while they "warm up" in the environment where the reader is before reading, so that the magnetic elements have time to stabilize at normal temperature.

How to get a LTO tape drive and system running

To get started with LTO tape drives and which one to choose and how to make your own rig, Matthew Millman made an excellent tutorial on which we build upon below, so you should read this tutorial and then read the instructions below.

The process is as follows: first find a second-hand/refurbished LTO drive with the highest revision you can for your budget, then find a server of a similar generation, or make an eGPU + SAS card of the highest speed the tape drive can support. Generally, you can aim for a LTO drive 3-4 generations older than the latest one (eg, if current is LTO9, you can expect cheap - 150-300 dollars per drive) for a LTO5 or LTO6). Aim only for LTO5+, because only LTFS did not exist before LTO5, but keep in mind some LTO5 drives need a firmware update to support LTFS, whereas all LTO6 drives support out of the box.

Once you find a second-hand LTO drive, consult its user manual beforehand to see what SAS or fibre cable (FC) you need (if SAS, any version should work, even greater versions, but older versions will just limit the read/write speed performance). For example, here is the manual for the HP LTO6 drive. All LTO drives are compatible with all computers provided you have the adequate connectivity (a SAS or FC adapter).

Once you have a LTO drive, then you can look for a computer to plug your LTO to. Essentially, you just need a computer that supports SAS. If not, then at least a free PCIe or mini-PCIe slot to be able to connect a SAS adapter.

The general outline is that you just need to have a computer with a PCIe slot, and get a SAS or FC adapter (depending on whether your LTO drive is SAS or FC) so that you can plug your LTO drive. There is currently no SAS to USB adapter, and only one manufacturer makes LTO drives with USB ports but they are super expensive, so just stick with internal SAS or FC drives (usually you want SAS, FC are better for long range connections, whereas SAS is compatible with SATA and SCSI drives, so you can also plug all your other hard drives plus the LTO tape drive on the same SAS adapter with this protocol).

In practice, there are 2 different available cost-effective approaches:

• If you have an external tape drive, then the best is to get a (second-hand) eGPU casing, and a PCIe SAS adapter, that you will plug in the eGPU casing instead of a GPU card. The eGPU casing should support Thunderbolt so this is how you will connect to the SAS and hence to your tape drive: you connect your laptop to the eGPU casing, and the eGPU casing to the external tape drive via the SAS adapter in the eGPU casing. This usually costs about 150-200 euros/dollars as of 2023.

• An alternative is to buy a low footprint PCIe dock such as EXP GDC produces, which essentially replaces the eGPU casing. The disadvantage is that your PCIe SAS adapter will be exposed, but this can be more cost effective (especially in second hand, you can get them at 20-40 euros/dollars instead of 120-150 euros/dollars brand new). But remember you also need to buy a power supply unit!

• If you got an internal tape drive, which are usually cheaper than external ones, then the approach is different: instead of configuring a sort of SAS-to-Thunderbolt bridge, here you get a standalone computer with either a motherboard that natively supports SAS (which is usually the case of computers meant to be servers), or at least a motherboard with a PCIe slot to buy separately a PCIe SAS adapter, and you plug your internal drive inside. So you will not be able to connect your laptop directly to the tape drive, you will have to pilot the server (which is just a standard desktop computer). Given these requirements, you can either make such a server yourself, but then keep in mind you have to build the whole computer, with a motherboard, a power supply, RAM, CPU, network, etc. Or, the easiest and usually cheapest route, is to just buy an old server with SAS hard drives second-hand (and every other components already in it), of a similar or later generation than your tape drive. Indeed, if the server has SAS hard drives, then it means you can connect your SAS tape drive too, no need for an adapter! Usually you can get them for cheap, for example if you get a 3-4 previous gen tape drive (eg, LTO-6 when current is LTO-9), then you can easily get a server computer of a similar generation for 100-250 euros/dollars, and everything is ready for you. Just make sure not to get a rack/blade computer, get one in tower form, easier to manipulate. Search on second hand websites: "server sas", then check that the SAS speed is on par with what your tape drive can accept, but if lower or higher, no biggie, it will just be slower, but it should work nevertheless. May also have to buy the right connectors but not an issue, just check the manual of your tape drive. Note: avoid HP Enterprise (HPE) servers, as there is a suspicion of programmed obsolescence in the Smart Array's Smart Storage Battery.

The consumables, the tapes, can also be easily found second-hand and usually are very cheap, eg, LTO6 tapes are sold at 10-20 euros/dollars one, for a storage space of 3TB to 6.25TB per tape.

With both approaches, expect at the cheapest a total cost of about 500 euros/dollars for the tape drive and attachment system (eGPU casing or dedicated server) as of 2023, which is very good and amortizable very fast with just a few tapes, even compared to the cheapest hard drives!

A modern data curation strategy for individuals

Here is an example curation strategy, which is accessible to individuals and not just big data centers:

• Get a LTO>=5 drive. Essentially, the idea with LTO is that you can just dump a copy of your whole hard drives, since the cartridges are big and inexpensive. And you can regularly reformat and overwrite the previous copy with a newer one. Store some LTO cartridges out of side to be robust against fires.
• If you want additional protection, especially by adding error-correction codes, DAR can be used to compress the data with PAR2 and is compatible with LTO. Alternatively, pyFileFixity can also be used to generate ECC codes, that can either be stored on the same cartridge alongside the files or on a separate cartridge depending on your threat model.
• Two kinds of archival plans are possible:

1. either only use LTO cartridges, then try to use cartridges of different brands (to avoid them failing at the same time - cartridges produced by the same industrial line will tend to include the same defects and similar lifespan) and store your data on at least 3 different copies/cartridges, per the redundancy principle (ie, "either bring one compass or three, but never two, because you will never know which one is correct").
2. either use LTO cartridges as ONE archival medium, and use other kinds of storage for the additional 2 copies you need: one can be an external hard drive, and the last one a cloud backup solution such as SpiderOak. The advantage of this solution is that it is more convenient: use your external hard drive to frequently backup, then also use your cloud backup to auto backup your most critical data online (off-site), and finally from time to time update your last copy on a LTO cartridge by mirroring your external hard drive.

• Curation strategy is then the same for all plans:

1. Every 5 years, the "small checkup": check your 3 copies, either by scanning sectors or by your own precomputed hashes (pyFileFixity's hash command).
2. If there is an error, assume the whole medium is dead and needs to be replaced and your data needs to be recovered: first using your error correction codes if you have, and then using pyFileFixity dup command to use a majority vote to reconstruct one valid copy out of the 3 copies.
3. Every 10 years, the "big checkup": even if the mediums did not fail, replace them by newer ones: mirror the old hard drive to a new one, the old LTO cartridge to a new one (it can be on a newer LTO version, so that you keep pace with the technology), etc.

With the above strategy, you should be able to preserve your data for as long as you can actively curate it. In case you want more robustness against accidents or the risk that 2 copies get corrupted under 5 years, then you can make more copies, preferably as LTO cartridges, but it can be other hard drives.

For more information on how to cold store LTO drives, read pp32-33 "Caring for Cartridges" instruction of this user manual. For HP LTO6 drives, Matthew Millman made an open-source commandline tool to do advanced LTO manipulations on Windows: ltfscmd.

In case you cannot afford a LTO drive, you can replace these by external hard drives, as they are less expensive to start with, but then your curation strategy should be done more frequently (ie, every 2-3 years a small checkup, and every 5 years, a big checkup).

Tools like pyFileFixity (or which can be used as complements)

Here are some tools with a similar philosophy to pyFileFixity, which you can use if they better fit your needs, either as a replacement of pyFileFixity or as a complement (pyFileFixity can always be used to generate an ecc file):

• DAR (Disk ARchive): similar to tar but non-solid thus allows for partial recovery and per-file access, plus it saves the directory tree meta-data -- see catalog isolation -- plus it can handle error correction natively using PAR2 and encryption. Also supports incremental backup, thus it's a very nice versatile tool. Crossplatform and opensource. Compatible with Linear Tape Open (LTO) magnetic bands storage (see instructions here)
• DVDisaster: error correction at the bit level for optical mediums (CD, DVD and BD / BluRay Discs). Very good, it also protects directory tree meta-data and is resilient to corruption (v2 still has some critical spots but v3 won't have any).
• rsbep tool that is part of dvbackup package in Debian: allows to generate an ecc of a stream of bytes. Great to pipe to dar and/or gz for your backups, if you're on unix or using cygwin.
• rsbep modification by Thanassis Tsiodras: enhanced rsbep to avoid critical spots and faster speed. Also includes a "freeze" script to encode your files into a virtual filesystem (using Python/FUSE) so that even meta-data such as directory tree are fully protected by the ecc. Great script, but not maintained, it needs some intensive testing by someone knowledgeable to guarantee this script is reliable enough for production.
• Parchive (PAR1, PAR2, MultiPar): well known error correction file generator. The big advantage of Parchives is that an ecc block depends on multiple files: this allows to completely reconstruct a missing file from scratch using files that are still available. Works good for most people, but most available Parchive generators are not satisfiable for me because 1- they do not allow to generate an ecc for a directory tree recursively (except MultiPar, and even if it is allowed in the PAR2 specs), 2- they can be very slow to generate (even with multiprocessor extensions, because the galois field is over 2^16 instead of 2^8, which is very costly), 3- the spec is not very resilient to errors and tampering over the ecc file, as it assumes the ecc file won't be corrupted (I also tested, it's still a bit resilient, but it could be a lot more with some tweaking of the spec), 4- it doesn't allow for partial recovery (recovering blocks that we can and pass the others that are unrecoverable): with PAR2, a file can be restored fully or it cannot be at all.
• Zip (with DEFLATE algorithm, using 7-Zip or other tools): allows to create non-solid archives which are readable by most computers (ubiquitous algorithm). Non-solid archive means that a zip file can still unzip correct files even if it is corrupted, because files are encoded in blocks, and thus even if some blocks are corrupted, the decoding can happen. A fast implementation with enhanced compression is available in pure Go (good for long storage).
• TestDisk: for file scraping, when nothing else worked.
• dd_rescue: for disk scraping (allows to forcefully read a whole disk at the bit level and copy everything it can, passing bad sector with options to retry them later on after a first full pass over the correct sectors).
• ZFS: a file system which includes ecc correction directly. The whole filesystem, including directory tree meta-data, are protected. If you want ecc protection on your computer for all your files, this is the way to go.
• Encryption: technically, you can encrypt your files without losing too much redundancy, as long as you use an encryption scheme that is block-based such as DES: if one block gets corrupted, it won't be decryptable, but the rest of the files' encrypted blocks should be decryptable without any problem. So encrypting with such algorithms leads to similar files as non-solid archives such as deflate zip. Of course, for very long term storage, it's better to avoid encryption and compression (because you raise the information contained in a single block of data, thus if you lose one block, you lose more data), but if it's really necessary to you, you can still maintain high chances of recovering your files by using block-based encryption/compression (note: block-based encryption can be seen as the equivalent of non-solid archives for compression, because the data is compressed/encrypted in independent blocks, thus allowing partial uncompression/decryption).
• SnapRAID
• par2ools: a set of additional tools to manage par2 archives
• Checkm: a tool similar to rfigc.py
• BagIt with two python implementations here and here: this is a file packaging format for sharing and storing archives for long term preservation, it just formalizes a few common procedures and meta data that are usually added to files for long term archival (such as MD5 digest).
• RSArmor a tool based on Reed-Solomon to encode binary data files into hexadecimal, so that you can print the characters on paper. May be interesting for small datasets (below 100 MB).
• Ent a tool to analyze the entropy of your files. Can be very interesting to optimize the error correction algorithm, or your compression tools.
• HashFS is a non-redundant, duplication free filesystem, in Python. Data deduplication is very important for large scale long term storage: since you want your data to be redundant, this means you will use an additional storage space for your redundant copies that will be proportional to your original data. Having duplicated data will consume more storage and more processing time, for no benefit. That's why it's a good idea to deduplicate your data prior to create redundant copies: this will be faster and save you money. Deduplication can either be done manually (by using duplicates removers) or systematically and automatically using specific filesystems such as zfs (with deduplication enabled) or hashfs.
• Paper as a storage medium: paper is not a great storage medium, because it has low storage density (ie, you can only store at most about 100 KB) and it can also degrade just like other storage mediums, but you cannot check that automatically since it's not digital. However, if you are interested, here are a few softwares that do that: Paper key, Paperbak, Optar, dpaper, QR Backup, QR Backup (another), QR Backup (again another), QR Backup (again), and finally a related paper.
• AVPreserve tools, most notably fixity to monitor for file changes (similarly to rfigc, but actively as a daemon) and interstitial to detect interstitial errors in audio digitization workflows (great to ensure you correctly digitized a whole audio file into WAV without any error).

FAQ

• Can I compress my data files and my ecc file?

As a rule of thumb, you should ALWAYS keep your ecc file in clear text, so under no compression nor encryption. This is because in case the ecc file gets corrupted, if compressed/encrypted, the decompression/decrypting of the corrupted parts may completely flaw the whole structure of the ecc file.

Your data files, that you want to protect, should remain in clear text, but you may choose to compress them if it drastically reduces the size of your files, and if you raise the resilience rate of your ecc file (so compression may be a good option if you have an opportunity to trade the file size reduction for more ecc file resilience). Also, make sure to choose a non-solid compression algorithm like DEFLATE (zip) so that you can still decode correct parts even if some are corrupted (else with a solid archive, if one byte is corrupted, the whole archive may become unreadable).

However, in the case that you compress your files, you should generate the ecc file only after compression, so that the ecc file applies to the compressed archive instead of the uncompressed files, else you risk being unable to correct your files because the uncompression of corrupted parts may output gibberish, and length extended corrupted parts (and if the size is different, Reed-Solomon will just freak out).

• Can I encrypt my data files and my ecc file ?

NEVER encrypt your ecc file, this is totally useless and counterproductive.

You can encrypt your data files, but choose a non-solid algorithm (like AES if I'm not mistaken) so that corrupted parts do not prevent the decoding of subsequent correct parts. Of course, you're lowering a bit your chances of recovering your data files by encrypting them (the best chance to keep data for the long term is to keep them in clear text), but if it's really necessary, using a non-solid encrypting scheme is a good compromise.

You can generate an ecc file on your encrypted data files, thus after encryption, and keep the ecc file in clear text (never encrypt nor compress it). This is not a security risk at all since the ecc file does not give any information on the content inside your encrypted files, but rather just redundant info to correct corrupted bytes (however if you generate the ecc file on the data files before encryption, then it's clearly a security risk, and someone could recover your data without your permission).

• What medium should I use to store my data?

The details are long and a bit complicated (I may write a complete article about it in the future), but the tl;dr answer is that you should use optical disks, because it decouples the storage medium and the reading hardware (eg, at the opposite we have hard drives, which contains both the reading hardware and the storage medium, so if one fails, you lose both) and because it's most likely future-proof (you only need a laser, which is universal, the laser's parameters can always be tweaked).

From scientific studies, it seems that, at the time of writing this (2015), BluRay HTL disks are the most resilient against environmental degradation. To raise the duration, you can also put optical disks in completely opaque boxes (to avoid light degradation) and in addition you can put any storage medium (not only optical disks, but also hard drives and anything really) in completely air-tight and water-tight bags or box and put in a fridge or a freezer. This is a law of nature: lower the temperature, lower will be the entropy, in other words lower will be the degradation over time. It works the same with digital data.

• What file formats are the most recoverable?

It's difficult to advise a specific format. What we can do is advise the characteristics of a good file format:

• future-proof (should be readable in the future).
• non-solid (ie, divised into indepedent blocks, so that a corruption to one block doesn't cause a problem to the decoding of other blocks).
• open source implementation available.
• minimize corruption impact (ie, how much of the file becomes unreadable with a partial corruption? Only the partially corrupted area, or other valid parts too?).
• No magic bytes or header importance (ie, corrupting the header won't prevent opening the file).

There are a few studies about the most resilient file formats, such as:

• "Just one bit in a million: On the effects of data corruption in files" by Volker Heydegger.
• "Analysing the impact of file formats on data integrity" by Volker Heydegger.
• "A guide to formats", by The UK national archives (you want to look at the Recoverability entry in each table).

• What is Reed-Solomon?

If you have any question about Reed-Solomon codes, the best place to ask is probably here (with the incredible Dilip Sarwate): http://www.dsprelated.com/groups/comp.dsp/1.php?searchfor=reed%20solomon

Also, you may want to read the following resources:

• "Reed-Solomon codes for coders", free practical beginner's tutorial with Python code examples on WikiVersity. Partially written by one of the authors of the present software.
• "Algebraic codes for data transmission", Blahut, Richard E., 2003, Cambridge university press. Readable online on Google Books.