3.2. Data integrity

So far, we mastered quite a number of challenges: Creating and populating a dataset with large and small files, modifying content and saving the changes to history, installing datasets, even as subdatasets within datasets, recording the impact of commands on a dataset with the run and re-run commands, and capturing plenty of provenance on the way. We further noticed that when we modified content in notes.txt or list_files.py, the modified content was in a text file. We learned that this precise type of file, in conjunction with the initial configuration template text2git we gave to datalad create, is meaningful: As the text file is stored in Git and not git-annex, no content unlocking is necessary. As we saw within the demonstrations of datalad run, modifying content of non-text files, such as .jpgs, requires – spoiler: at least in our current type of dataset and if you are not on Windows – the additional step of unlocking file content, either by hand with the datalad unlock command, or within datalad run using the -o/--output flag.

There is one detail about DataLad datasets that we have not covered yet. Its both a crucial aspect to understanding certain aspects of a dataset, but it is also a potential source of confusion that we want to eradicate.

You might have noticed already that an ls -l or tree command in your dataset shows small arrows and quite cryptic paths following each non-text file. Maybe your shell also displays these files in a different color than text files when listing them. We’ll take a look together, using the books/ directory as an example:

This will look different to you

First of all, the tree equivalent provided by condas m2-base package doesn’t list individual files, only directories. And, secondly, even if you list the individual files (e.g., with ls -l), you would not see the symlinks shown below. Due to insufficient support of symlinks on Windows, git-annex does not use them. Please read on for a basic understanding of how git-annex usually works – a Windows-Workaround at the end of this section will then highlight the difference in functionality on Windows.

# in the root of DataLad-101
$ cd books
$ tree
.
├── bash_guide.pdf -> ../.git/annex/objects/WF/Gq/MD5E-s1198170--0ab2c121bcf68d7278af266f6a399c5f.pdf/MD5E-s1198170--0ab2c121bcf68d7278af266f6a399c5f.pdf
├── byte-of-python.pdf -> ../.git/annex/objects/z1/Q8/MD5E-s4208954--ab3a8c2f6b76b18b43c5949e0661e266.pdf/MD5E-s4208954--ab3a8c2f6b76b18b43c5949e0661e266.pdf
├── progit.pdf -> ../.git/annex/objects/G6/Gj/MD5E-s12465653--05cd7ed561d108c9bcf96022bc78a92c.pdf/MD5E-s12465653--05cd7ed561d108c9bcf96022bc78a92c.pdf
└── TLCL.pdf -> ../.git/annex/objects/jf/3M/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf

0 directories, 4 files

If you do not know what you are looking at, this looks weird, if not worse: intimidating, wrong, or broken. First of all: no, it is all fine. But let’s start with the basics of what is displayed here to understand it.

The small -> symbol connecting one path (the book’s name) to another path (the weird sequence of characters ending in .pdf) is what is called a symbolic link (short: symlink) or softlink. It is a term for any file that contains a reference to another file or directory as a relative path or absolute path. If you use Windows, you are familiar with a related, although more basic concept: a shortcut.

This means that the files that are in the locations in which you saved content and are named as you named your files (e.g., TLCL.pdf), do not actually contain your files’ content: they just point to the place where the actual file content resides.

This sounds weird, and like an unnecessary complication of things. But we will get to why this is relevant and useful shortly. First, however, where exactly are the contents of the files you created or saved?

The start of the link path is ../.git. The section Create a dataset contained a note that strongly advised that you to not temper with (or in the worst case, delete) the .git repository in the root of any dataset. One reason why you should not do this is because this .git directory is where all of your file content is actually stored.

But why is that? We have to talk a bit git-annex now in order to understand it1.

When a file is saved into a dataset to be tracked, by default – that is in a dataset created without any configuration template – DataLad gives this file to git-annex. Exceptions to this behavior can be defined based on

  1. file size

  2. and/or path/pattern, and thus for example file extensions, or names, or file types (e.g., text files, as with the text2git configuration template).

git-annex, in order to version control the data, takes the file content and moves it under .git/annex/objects – the so called object-tree. It further renames the file into the sequence of characters you can see in the path, and in its place creates a symlink with the original file name, pointing to the new location. This process is often referred to as a file being annexed, and the object tree is also known as the annex of a dataset.

What happens on Windows?

Windows has insufficient support for symlinks and revoking write permissions on files. Therefore, git-annex classifies it as a crippled filesystem and has to stray from its default behavior. While git-annex on Unix-based file operating systems stores data in the annex and creates a symlink in the data’s original place, on Windows it moves data into the annex and creates a copy of the data in its original place.

Why is that? Data needs to be in the annex for version control and transport logistics – the annex is able to store all previous versions of the data, and manage the transport to other storage locations if you want to publish your dataset. But as the Findoutmore at the end of this section will show, the annex is a non-human readable tree structure, and data thus also needs to exist in its original location. Thus, it exists in both places: its moved into the annex, and copied back into its original location. Once you edit an annexed file, the most recent version of the file is available in its original location, and past versions are stored and readily available in the annex. If you reset your dataset to a previous state (as is shown in the section Back and forth in time), the respective version of your data is taken from the annex and copied to replace the newer version, and vice versa.

But doesn’t a copy mean data duplication? Yes, absolutely! And that is a big downside to DataLad and git-annex on Windows. If you have a dataset with annexed file contents (be that a dataset you created and populated yourself, or one that you cloned and got file contents with datalad get from), it will take up more space than on a Unix-based system. How much more? Every file that exists in your file hierarchy exists twice. A fresh dataset with one version of each file is thus twice as big as it would be on a Linux computer. Any past version of data does not exist in duplication.

Step-by-step demonstration: Let’s take a concrete example to explain the last point in more detail. How much space, do you think, is taken up in your dataset by the resized salt_logo_small.jpg image? As a reminder: It exists in two versions, a 400 by 400 pixel version (about 250Kb in size), and a 450 by 450 pixel version (about 310Kb in size). The 400 by 400 pixel version is the most recent one. The answer is: about 810Kb (~0.1Mb). The most recent 400x400px version exists twice (in the annex and as a copy), and the 450x450px copy exists once in the annex. If you would reset your dataset to the state when we created the 450x450px version, this file would instead exist twice.

Can I at least get unused or irrelevant data out of the dataset? Yes, either with convenience commands (e.g., git annex unused followed by git annex dropunused), or by explicitly using drop on files (or there past versions) that you don’t want to keep anymore. Alternatively, you can transfer data you don’t need but want to preserve to a different storage location. Later parts of the handbook will demonstrate each of these alternatives.

For a demonstration that this file path is not complete gibberish, take the target path of any of the book’s symlinks and open it, for example with evince <path> (Note: exchange evince with your standard PDF reader).

evince ../.git/annex/objects/jf/3M/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf

Even though the path looks cryptic, it works and opens the file. Whenever you use a command like evince TLCL.pdf, internally, your shell will follow the same cryptic symlink like the one you have just opened.

But why does this symlink-ing happen? Up until now, it still seems like a very unnecessary, superfluous thing to do, right?

The resulting symlinks that look like your files but only point to the actual content in .git/annex/objects are small in size. An ls -lah reveals that all of these symlinks have roughly the same, small size of ~130 Bytes:

$ ls -lah
total 24K
drwxr-xr-x 2 adina adina 4.0K Jan 29 08:37 .
drwxr-xr-x 7 adina adina 4.0K Jan 29 08:38 ..
lrwxrwxrwx 1 adina adina  131 Aug 23 12:45 bash_guide.pdf -> ../.git/annex/objects/WF/Gq/MD5E-s1198170--0ab2c121bcf68d7278af266f6a399c5f.pdf/MD5E-s1198170--0ab2c121bcf68d7278af266f6a399c5f.pdf
lrwxrwxrwx 1 adina adina  131 Jun 16  2020 byte-of-python.pdf -> ../.git/annex/objects/z1/Q8/MD5E-s4208954--ab3a8c2f6b76b18b43c5949e0661e266.pdf/MD5E-s4208954--ab3a8c2f6b76b18b43c5949e0661e266.pdf
lrwxrwxrwx 1 adina adina  133 Jun 29  2019 progit.pdf -> ../.git/annex/objects/G6/Gj/MD5E-s12465653--05cd7ed561d108c9bcf96022bc78a92c.pdf/MD5E-s12465653--05cd7ed561d108c9bcf96022bc78a92c.pdf
lrwxrwxrwx 1 adina adina  131 Jan 28  2019 TLCL.pdf -> ../.git/annex/objects/jf/3M/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf

Here you can see the reason why content is symlinked: Small file size means that Git can handle those symlinks! Therefore, instead of large file content, only the symlinks are committed into Git, and the Git repository thus stays lean. Simultaneously, still, all files stored in Git as symlinks can point to arbitrarily large files in the object tree. Within the object tree, git-annex handles file content tracking, and is busy creating and maintaining appropriate symlinks so that your data can be version controlled just as any text file.

This comes with two very important advantages:

One, should you have copies of the same data in different places of your dataset, the symlinks of these files point to the same place (in order to understand why this is the case, you will need to read the hidden section at the end of the page). Therefore, any amount of copies of a piece of data is only one single piece of data in your object tree. This, depending on how much identical file content lies in different parts of your dataset, can save you much disk space and time.

The second advantage is less intuitive but clear for users familiar with Git.

Note for Git users

Small symlinks can be written very very fast when switching branches, as opposed to copying and deleting huge data files.

This leads to a few conclusions:

The first is that you should not be worried to see cryptic looking symlinks in your repository – this is how it should look. If you are interested in why these paths look so weird, and what all of this has to do with data integrity, you can check out the hidden section below.

The second is that it should now be clear to you why the .git directory should not be deleted or in any way modified by hand. This place is where your data are stored, and you can trust git-annex to be better able to work with the paths in the object tree than you or any other human are.

Lastly, understanding that annexed files in your dataset are symlinked will be helpful to understand how common file system operations such as moving, renaming, or copying content translate to dataset modifications in certain situations. Later in this book we will have a section on how to manage the file system in a DataLad dataset (Miscellaneous file system operations).

more about paths, checksums, object trees, and data integrity

But why does the target path to the object tree needs to be so cryptic? Does someone want to create maximal confusion with this naming? Can’t it be … more readable?

Its not malicious intent that leads to these paths and file names. Its checksums. And they are quite readable – just not for humans, but git-annex. Understanding the next section is completely irrelevant for the subsequent sections of the book. But it can help to establish trust in that your data are safely stored and tracked, and it can get certainly helpful should you be one of those people that always want to understand things in depth. Also, certain file management operations can be messy – for example, when you attempt to move a subdirectory (more on this in a dedicated section Miscellaneous file system operations) it can break symlinks, and you need to take appropriate actions to get the dataset back into a clean state. Understanding more about the object tree can help to understand such problems, and knowing bits of the git-annex basics can make you more confident in working with your datasets.

So how do these paths and names come into existence?

When a file is annexed, git-annex generates a key from the file content. It uses this key (in part) as a name for the file and as the path in the object tree. Thus, the key is associated with the content of the file (the value), and therefore, using this key, file content can be identified – or rather: Based on the keys, it can be identified whether two files have identical contents, and whether file content changed.

The key is generated using hashes. A hash is a function that turns an input (e.g., a PDF file) into a string of characters with a fixed length. In principle, therefore, the hash function simply transforms a content of any size into a string with fixed length.

The important aspect of a hash function is that it will generate the same hash for the same file content, but once file content changes, the generated hash will also look different. If two files are turned into identical character strings, the content in these files is thus identical. Therefore, if two files have the same symlink, and thus link the same file in the object-tree, they are identical in content. If you have many copies of the same data in your dataset, the object tree will contain only one instance of that content, and all copies will symlink to it, thus saving disk space. But furthermore, the file name also becomes a way of ensuring data integrity. File content can not be changed without git-annex noticing, because the symlink to the file content will change. If you want to read more about the computer science basics about about hashes check out the Wikipedia page here.

This key (or checksum) is the last part of the name of the file the symlink links to (in which the actual data content is stored). The extension (e.g., .pdf) is appended because some operating systems (Windows) need this information. The key is also one of the subdirectory names in the path. This subdirectory adds an important feature to the object-tree: It revokes the users permissions to modify it. This two-level structure is implemented because it helps to prevent accidental deletions and changes, and this information will be helpful to understand some file system management operations (see section Miscellaneous file system operations), for example deleting a subdataset.

# take a look at the last part of the target path:
$ ls -lah TLCL.pdf
lrwxrwxrwx 1 adina adina 131 Jan 28  2019 TLCL.pdf -> ../.git/annex/objects/jf/3M/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf/MD5E-s2120211--06d1efcb05bb2c55cd039dab3fb28455.pdf
# compare it to the checksum (here of type md5sum) of the PDF file and the subdirectory name
$ md5sum TLCL.pdf
06d1efcb05bb2c55cd039dab3fb28455  TLCL.pdf

There are different hash functions available. Depending on which is used, the resulting checksum has a certain length and structure. By default, DataLad uses MD5E checksums, but should you want to, you can change this default to one of many other types. The first part of the file name actually states which hash function is used. The reason why MD5E is used is because it is comparatively short – thus it is possible to share your datasets also with users on operating systems that have restrictions on total path lengths (Windows). Therefore, refrain from changing this default if you are on Windows, or want Windows user to be able to use your dataset.

By now we know where almost all parts of the file name derived from – the remaining unidentified bit in the file name is the one after the checksum identifier. This part is the size of the content in bytes. An annexed file in the object tree thus has a file name following this structure:

checksum-identifier - size -- checksum . extension

As a last puzzle piece to shed some light onto the path in the object tree, there are two more directories on top of the subdirectory named after the checksum, just after .git/annex/objects/, consisting of two letters each. These two letters are also derived from the md5sum of the key, and their sole purpose to exist is to avoid issues with too many files in one directory (which is a situation that certain file systems have problems with).

In summary, you now know a great deal about git-annex and the object tree. Maybe you are as amazed as we are about some of the ingenuity used behind the scenes. In any case, this section was hopefully insightful, and not confusing. If you are still curious about git-annex, you can check out its documentation.