Under construction

The main reference for this tutorial is the Pro Git book section on GIT internals.

This tutorial uses three libraries:

CodeMirror, released under the MIT license
sha1.js, released under the MIT license
pako 2.0.3, released under the MIT and Zlib licenses, see the project page for details.
Viz.js (v1.8.2 which has a synchronous API), released under the MIT license

Introduction

GIT is based on a simple model, with a lot of shorthands for common use cases. This model is sometimes hard to guess just from the everyday commands. To illustrate how GIT works, we'll implement a stripped down clone of GIT in a few lines of JavaScript. * empty lines and single closing braces excluded, a few more in total.

The Operating System's filesystem

Model of the filesystem

The Operating System's filesystem will be simulated by a very simple key-value store. In this very simple filesystem, directories are entries mapped to null and files are entries mapped to strings. The path to the current directory is stored in a separate variable.

Filesystem access functions (`read`, `write`, `mkdir`, `exists`, `remove`, `cd`)

The filesystem exposes functions to read an entire file, create or replace an entire file, create a directory, test the existence of a filesystem entry, and change the current directory.

Filesystem access functions (`listdir`)

It will be handy for some operations to list the contents of a directory.

Example working directory

Our imaginary user will create a proj directory, and start filling in some files.

`git init` (creating `.git`)

The first thing to do is to initialize the GIT directory. For now, only the .git folder is needed, The rest of the function implementing git init will be written later.

Click on the eval button to see the files and directories that were created so far.

`git hash-object` (storing a copy of a file in `.git`)

The most basic element of a GIT repository is an object. Objects have a type which can be blob (individual files), tree (directories), commit (pointers to a specific version of the root directory, with a description and some metadata) and tag (named pointers to a specific commit, with a description and some metadata). When a file is added to the git repostitory, a compressed copy is stored in GIT's database, in the .git/objects/ folder. This copy is a blob object.

The compressed copy is given a unique filename, which is obtained by hashing the contents of the original file. Some filesystems have poor performance when a single directory contains a large number of files, and some filesystems have a limit on the number of files that a directory may contain. To circumvent these issues, the first two characters of the hash are used as the name of an intermediate directory: if a file's hash is 0a1bd…, its compressed copy will be stored in .git/objects/0a/1bd…

This function creates a file that looks like this:

The objects stored in the GIT database are compressed with zlib (using the "deflate" compression method). The filesystem view shows the marker deflated: followed by the uncompressed data. Click on the (un)compressed data to toggle between this pretty-printed view and the raw compressed data.

When creating some blob objects, the result could be, for example:

This function reproduces faithfully the behaviour of (a subset of the options of) the git hash-object command which can be called on a real git command-line.

Adding a file to the GIT database

So far, our GIT database does not know about any of the user's files. In order to add the contents of the README file in the database, we use git hash-object -w -t blob README, where -w tells GIT to write the object in its database, and -t blob indicates that we want to create a blob object, i.e. the contents of a file.

Click on the eval button to see the file that was created by this call.

You can notice that the database does not contain the name of the original file, only its content, stored under a unique identifier which is derived by hashing that content. Let's add the second user file to the database.

`zlib` compression

GIT compresses objects with zlib. The deflate() function used in the script above comes from the pako 2.0.3 library. To view a zlib-compressed object in your *nix terminal, simply write this declaration in your shell.

unzlib() {
  python -c \
    "import sys,zlib; \
     sys.stdout.buffer.write(zlib.decompress(open(sys.argv[1], 'rb').read()));" \
    "$1"
}

You can then inspect git objects as follows, using hexdump to view the null bytes and other non-printable bytes.

unzlib .git/objects/95/d318ae78cee607a77c453ead4db344fc1221b7 | hexdump -Cv

Storing trees (list of hashed files and subtrees)

At this point GIT knows about the contents of both of the user's files, but it would be nice to also store the filenames. This is done by creating a tree object

A tree object can contain files (by associating the blob's hash to its name), or directories (by associating the hash of other subtrees to their name). The mode (100644 for the file and 40000 for the folder) incidates the permissions, and is given in octal using the values used by *nix

In the contents of a tree, subdirectories (trees) are listed before files (blobs); within each group the entries are ordered alphabetically.

This function needs a small utility to convert hashes encoded in hexadecimal to raw bytes.

Example use of `store_tree()`

The following code, once uncommented, stores into the GIT database the trees for src and for the root directory of the GIT project.

The store_tree() function needs to be called for the contents of subdirectories first, and that result can be used to store the trees of upper directories. In the next section, we will write a function which takes a list of paths, constructs an internal representation of the hierarchy, and stores the corresponding trees bottom-up.

Storing a tree from a list of paths

Making trees out of the subfolders one by one is cumbersome. The following utility function takes a list of paths, and builds a tree from those.

function paths_to_tree(paths) {
  // This temporary mutable object will store a hierarchy of
  // subfolders and files, e.g.
  // {
  //   subfolders: { src: { subfolders: [], files: ['main.scm'] } }
  //   files: ['README']
  // }
  var hierarchy = { subfolders: {}, files: [] };

// This splits the input paths on occurrences of "/",
  // and inserts them into the "hierarchy" object.
  for (var i = 0; i < paths.length; i++) {
    var path_components = paths[i].split('/');
    var h = hierarchy;
    for (var j = 0; j < path_components.length - 1; j++) {
      if (! h.subfolders.hasOwnProperty(path_components[j])) {
        h.subfolders[path_components[j]] = {
          subfolders: {},
          files: []
        };
      }
      h = h.subfolders[path_components[j]];
    }
    h.files[h.files.length] = path_components[path_components.length - 1];
  }

// This function takes the path to a directory, e.g. "src",
  // and a hierarchy object e.g. { subfolders: [], files: ['main.scm'] }.
  // It recursively stores the tree object for that directory into
  // GIT's database.
  var to_tree = function(base_directory, hierarchy) {
    var subtrees = [];
    for (var i in hierarchy.subfolders) {
      if (hierarchy.subfolders.hasOwnProperty(i)) {
        subtrees[subtrees.length] = {
          name: i,
          hash: to_tree(join_paths(base_directory, i), hierarchy.subfolders[i])
        };
      }
    }
    return store_tree(base_directory, hierarchy.files, subtrees);
  }

// Store the trees for the whole hierarchy, starting from the
  // root directory of the GIT repository (which is represented
  // as an empty path "")
  return to_tree("", hierarchy);
}

// git add README src/main.scm
paths_to_tree(["README", "src/main.scm"]);

Storing a commit in the GIT database

Now that the GIT database contains the entire tree for the current version, a commit can be created. A commit contains

the hash of the tree object,
the hash of the previous commit, which is dubbed the parent (merge commits have two or more parents, and the initial commit has no parent commit),
information about the author (the person who initially wrote the code),
information about the committer (the person who adds the code to the GIT database, often the same person as the author, but it can be a different person e.g. when someone else rewrites the history with a rebase or applies a patch recieved by e-mail),
and a description.

The author and committer information contain

the person's name,
the person's email,
the *nix timestamp at which the version was authored or committed,
and the timezone for that timestamp.

function store_commit(tree, parents, author, committer, message) {
  var commit_contents = '';
  commit_contents += 'tree ' + tree + '\n';
  for (var i = 0; i < parents.length; i++) {
    commit_contents += 'parent ' + parents[i] + '\n';
  }
  commit_contents += 'author ' + author.name
                  + ' <' + author.email + '> '
                  + format_date(author.date) + ' '
                  + format_timezone(author.timezoneMinutes) + '\n';
  commit_contents += 'committer ' + committer.name
                  + ' <' + committer.email + '> '
                  + format_date(committer.date) + ' '
                  + format_timezone(committer.timezoneMinutes) + '\n';
  commit_contents += '\n';
  commit_contents += '' + message + (message[message.length-1] == '\n' ? '' : '\n');
  // cat commit_contents | git hash-object -w -t commit --stdin
  return hash_object(true, 'commit', true, commit_contents);
}

Storing an example commit

It is now possible to store a commit in the database. This saves a copy of the tree along with some metadata about this version. The first commit has no parent, which is represented by passing the empty list.

resolving references

The next few subsections will introduce symbolic references and other references like branch names, the special name HEAD or tag names.

Most GIT commands accept as an argument a commit hash or a named reference to a hash. In order to implement those, we need to be able to resolve these references first.

Symbolic references are nothing more than regular files containing a hexadecimal hash or a string of the form ref: path/to/other/symbolic/reference. The HEAD reference is stored in .git/HEAD, and can point directly to a commit hash like 0123456789abcdef0123456789abcdef01234567, or can point to another symbolic reference, in which case the .git/HEAD file will contain e.g. refs/heads/main.

Branches are simple files stored in .git/refs/heads/name-of-the-branch and usually contain a hash like 0123456789abcdef0123456789abcdef01234567.

Tags are identical to branches in terms of representation. It seems that the only difference between tags and branches is the behaviour of git checkout and similar commands. These commands, as explained in the section about git checkout below, normally write ref: refs/heads/name-of-branch in .git/HEAD when checking out a branch, but write the hash of the target commit when checking out a tag or any other non-branch reference.

We'll start with a small utility to remove the newline at the end of a string. GIT references are usually files containing a hexadecimal hash, and following *NIX tradition these files finish with a newline byte. When reading these references, we need to get rid of the newline first.

`git symbolic-ref`

git symbolic-ref is a low-level command which reads (and in the official GIT implementation also writes and updates) symbolic references given a path relative to .git/. For example, git symbolic-ref HEAD will read the contents of the file .git/HEAD, and if that file starts with ref: , the rest of the line will be returned.

The official implementation of GIT follows references recursively and returns the path/to/file of the last file of the form ref: path/to/file. In the example below, git symbolic-ref HEAD would

read the file proj/.git/HEAD which contains ref: refs/heads/main,
follow that indirection and read the file proj/.git/refs/heads/main which contains ref: refs/heads/other
follow that indirection and read the file proj/.git/refs/heads/other which contains a hash
return the last file path that contained a ref:, i.e. return the string refs/heads/other

`git rev-parse`

git rev-parse is another low-level command. It takes a symbolic reference or other reference, and returns the hash. The difference with git symbolic-ref is that symbolic-ref follows indirections to other references, and returns the last named reference in the chain of indirections, whereas rev-parse goes one step further and returns the hash pointed to by the last named reference.

`git branch`

A branch is a pointer to a commit, stored in a file in .git/refs/heads/name_of_the_branch. The branch can be overwritten with git branch -f. Also, as will be explained later, git commit can update the pointer of a branch.

When we call git branch main HEAD or equivalently git branch main 0123456789012345678901234567890123456789, a file containing that hash is created in .git/refs/heads/main. This file acts as a pointer to the branch, and this pointer can be read e.g. by git rev-parse.

After creating the branch, we show how the file .git/refs/heads/main can be overwritten using git branch -f

`HEAD`

The HEAD indicates the "current" commit. It is set at first as part of the git init routine.

Usually, the HEAD is a symbolic reference to a branch, i.e. the file .git/HEAD contains ref: refs/heads/name-of-branch. When checking out a commit by specifying its hash directly, or when checking out a non-branch reference, the file .git/HEAD contains the hash of the commit instead.

The state in which .git/HEAD contains a commit hash is called "detached HEAD", and often sounds alarming to people who have not encountered this before. As we will see in the following sections, the only difference between detached HEAD and the normal state is that git commit updates the branch to point to the new commit in the normal mode of operation. When the HEAD is detached, it does not point to a specific branch, and git commit updates the HEAD directly instead, overwriting it with the new commit hash.

Since the HEAD is supposed to be a transient pointer, it is easy to lose track of the hash of an important commit. For example, the following sequence of operations:

git checkout 0123456789abcdef0123456789abcdef01234567

touch new_file
git add new_file
git commit -m 'This is a commit adding a new file'

git checkout branch-of-feature-foobar

roughly means:

HEAD = 0123456789abcdef0123456789abcdef01234567
// overwrite the contents of the working directory with
// the contents of commit 0123456789abcdef0123456789abcdef01234567
checkout(0123456789abcdef0123456789abcdef01234567)

// create commit with the new file:
HEAD = commit(…)

// Checkout other branch
HEAD = git_rev_parse('branch-of-feature-foobar')

The hash of the new commit which is stored in HEAD on the second step is overwritten in the third step. In order to later retrieve that specific version with the precious new_file, one needs that hash. It would be possible to note down these hashes in a simple text file, but GIT offers a mechanism for that: branches. After all, branches are merely named text files containing the hash of the latest commit in that line of work.

The hash of a commit created with git commit does not only exist in the HEAD file (when in detached HEAD) or in the current branch file (normal mode). The official implementation of GIT keeps a log of the changes being made to the various references. .git/logs/HEAD contains a log of the hashes pointed to by .git/HEAD, and .git/logs/refs/heads/main contains a log of the hashes pointed to by .git/refs/heads/main, and the commands git reflog and git reflog main pretty-print these files.

There are a few more ways to find a lost commit hash, including a careful invocation of git fsck which checks that the files stored in .git/ are not corrupted, and that no reference (to another reference or a commit, tree or blob) points to a non-existing file. The git fsck --unreachable option tells this command to print all object hashes which are not pointed to indirectly by any named reference (so-called unreachable objects, which are well-formed but are not indirectly linked to from a branch or other kind of named pointer).

The reflog can be used to recover a lost hash but handling hashes manually like this is somewhat error-prone, and most new users are not aware of those features; for this reason GIT commands tend to display a warning when switching to a detached HEAD state.

git config

The official implementation of GIT stores the settings in various files (.git/config within a repository, ~/.gitconfig in the user's home folder, and several other places).

These files use a .ini syntax with key = value lines grouped under some [section] headings. The configuration above could be stored in ~/.gitconfig or .git/config using the following syntax:

[user]
name = Ada Lovelace
email = ada@analyti.cal

The $EDITOR variable is a traditional *NIX environment variable, and could e.g. be declared with EDITOR=nano in ~/.profile or ~/.bashrc.

`git commit`

The git commit command stores a commit (metadata and a pointer to a tree containing the files given on the command-line), and updates the HEAD or current branch to point to the new commit.

If the HEAD points to a commit hash, then git commit updates the HEAD to point to the new commit. Otherwise, when the HEAD points to a branch, then the target branch (represented by a file named .git/refs/heads/the_branch_name) is updated.

The official implementation of git commit makes use of the index. When a file is scheduled for the next commit using git add path/to/file, it is added to the index. The index is a representation of a collection of copies of files, which can efficiently be compared to the working directory. It uses a different representation, but its role is very similar to that of a tree object along with the subtrees and blob objects of individual files. When git commit is called without specifying any files, it creates a commit containing the version of the files stored in the index.

In this simplified implementation, we only support creating commits by specifying all the files that must be present in the commit (including unchanged files). This contrasts with the official implementation which would create a tree containing the files from the current HEAD, as well as the added, modified or deleted files specified by git add or specified directly on the git commit command-line.

`git tag`

Tags behave like branches, but are stored in .git/refs/tags/the_tag_name and a tag is not normally modified. Once created, it's supposed to always point to the same version.

GIT does offer a git tag -f existing-tag new-hash command, but using it should be a rare occurrence.

Intuitively, tags differ from branches in the following way: when checking out a branch, and a subsequent commit is made, the branch is updated to point to the new commit's hash. As we've seen in the implementation of git commit, the difference is actually in the contents of the .git/HEAD file. If it is a symbolic reference (generally a pointer to a branch), then the target of that reference is updated every time a new commit is created. If the .git/HEAD file contains the hash of a commit, then the .git/HEAD file itself is updated every time a new commit is created.

Therefore, tags and branches differ only in their usage and in the path under which they are stored (.git/refs/heads/name-of-the-branch vs. .git/refs/tags/name-of-the-tag). The file .git/HEAD is overwritten by git commit and git checkout. It is the latter command which will behave differently for tags and branches; git checkout branch-name turns the HEAD into a symbolic reference, whereas git checkout tag-name resolves the tag name to a commit hash, and writes that hash directly into .git/HEAD.

`git checkout`

The git checkout commit-hash-or-reference command modifies the HEAD to point to the given commit, and modifies the working directory to match the contents of the tree object pointed to by that commit.

Checkout, branches and other references

The HEAD does not normally point to a tag. Although nothing actually prevents writing ref: refs/tags/v1.0 into .git/HEAD, the GIT commands will not automatically do this. For example, git checkout tag-or-branch-or-hash will put a symbolic ref: in .git/HEAD only if the argument is a branch.

Checking out files

In order to replace the contents of the working directory with those of the given commit, we recursively compare the subtrees, deleting from the working directory the files or directories that are not present in the tree object, and overwriting the others.

The official implementation of GIT will record the diff between the current working directory and the current commit, and will re-apply these changes on top of the freshly checked-out commit. The official git checkout command will print warnings and refuse to proceed when these changes cannot be re-applied without conflict, encouraging the user to create a commit containing this updated version or to stash the changes (effectively creating a temporary commit containing this version, pointed to by .git/refs/stash). Our simple implementation will always overwrite the changes.

function checkout_files(hash) {
  var commit = parse_commit(hash);
  checkout_tree(current_directory, commit.tree);
}

Assert

The checkout_tree() function needs to read the commit, tree and blob objects from the .git/ folder. The following sections will introduce some parsers for these objects. The parsers will check that their input looks reasonably well-formed, using assert().

Reading compressed objects

Parsing tree objects

The parse_tree function above needs a small utility to convert hashes represented using raw bytes to a hexadecimal representation.

Parsing commit objects

assert(tree,      'commit lacks tree header');
  assert(author,    'commit lacks author header');
  assert(committer, 'commit lacks committer header');

return {
    tree: tree,
    parents: parents,
    author: author,
    committer: committer,
    message: message
  };
}

Parsing author and committer metadata

Example checkout

`git init`

The git init command creates the .git directory and points .git/HEAD to the default branch (a file which does not exist yet, as this branch does not contain any commit at this point).

The index

When adding files with git add, GIT does not immediately create a commit object. Instead, it adds the files to the index, which uses a binary format with lots of metadata. The mock filesystem used here lacks most of these pieces of information, so thr value 0 will be used for most fields. See this blog post for a more in-depth study of the index.

function store_index(paths) {
  var magic = 'DIRC' // DIRectory Cache
  var version = raw_bytes32(2);
  var entries = raw_bytes32(paths.length);
  var header = magic + version + entries;

index = header;

for (var i = 0; i < paths.length; i++) {
    var ctime  = raw_bytes64(0);
    var mtime  = raw_bytes64(0);
    var device = raw_bytes32(0);
    var inode  = raw_bytes32(0);
    // default permissions for files, in octal.
    var mode   = raw_bytes32(0100644);
    var uid    = raw_bytes32(0);
    var gid    = raw_bytes32(0);
    var size   = raw_bytes32(read(join_paths(current_directory, paths[i])).length);
    var hash   = hex_to_raw_bytes(hash_object(true, 'blob', false, paths[i]));
    // for this simple index, the flags (the 4 higher bits) are 0.
    assert(paths[i].length < 0xfff)
    var flags_and_file_path_length = raw_bytes16(paths[i].length)
    var file_path = paths[i] + '\0';
    entry = ctime + mtime + device + inode + mode + uid + gid + size
          + hash + flags_and_file_path_length + file_path;
    while (entry.length % 8 != 0) {
      // pad with null bytes to a multiple of 8 bytes (64-bits).
      entry += '\0';
    }

index += entry;
  }

index += hex_to_raw_bytes(sha1(index));

write(join_paths(current_directory, '.git/index'), index)
}

Playground

The implementation is now sufficiently complete to create a small repository.

By clicking on "Copy commands to recreate in *nix terminal.", it is possible to copy a series of mkdir … and printf … > … commands that, when executed, will recreate the virtual filesystem on a real system. The resulting folder is bit-compatible with the official git log, git status, git checkout etc. commands.

Suggested exercises

The reader willing to improve their grasp of GIT's mental model, and reduce their reliance on a few learned recipies, might be interested in the following warm-up exercises:

Inspect an existing repository, starting with cat .git/HEAD and using git cat-file -p some-hash to pretty-print an object given its hash.
Inspect an existing repository, starting with cat .git/HEAD and using the zlib decompression tool from the zlib compression section.
Run git init new-directory in a terminal, and create an initial single-file commit from scratch, using only git hash-object, printf and overwriting .git/HEAD. This will involve retracing the steps in this tutorial to create a blob object for the file, a tree object to be the directory containing just that file, and a commit object.
For a couple of weeks, only use the GIT commands commit, diff, checkout, merge, cherry-pick, log, clone, fetch and push remote hash-of-commit:refs/heads/name-of-the-branch. In particular, don't use rebase which is just a wrapper around a sequence of cherry-pick commands, don't use pull which is just a wrapper around fetch and merge, don't use git push as-is and instead explicitly give the name (origin) or URL of the remote, the hash of the commit to push, and the path that should be updated on the remote (git push while the main branch is checked out locally is equivalent to git push origin HEAD:refs/heads/main, where HEAD can be replaced by the actual hash of the commit).
Try not even using git cherry-pick or git diff a few times, instead make two copies the git directoy, check out the two different commits in each copy, and use the traditional *NIX commands diff and patch.

Conclusion

This article shows that a large part of the core of GIT can be re-implemented in a few source lines of code* (copy all the code). * empty lines and single closing braces excluded, a few more in total.

Some of the features which may appear mysterious at first sight (e.g. detached HEAD) should be clearer with the knowledge of how GIT works behind the scenes.

Furthermore, branches are often associated with an intuition (containers into which commits are added) which does not match the implementation (mutable pointers to commits).

Finally, it is tempting to think of commits as patches. While darcs tries to expose an interface which matches this intuition, it is clear that the implementation of GIT considers commits as copies of the entire repository, and are linked to the previous version solely by the parent metadata in the commit headers.

A few core commands like git diff and git apply are not described in this tutorial. They are little more than improved versions of the classical *nix commands diff and patch.

Most other commands provided by GIT are merely convenience wrappers around these commands. For example, git cherry-pick is simply a combination of git diff between the tree of a commit and the tree of its parent, followed by git apply to apply the patch and git commit to create a new commit whose diff is equivalent to the diff of the original commit. As an other example, the command git rebase performs as succession of cherry-pick operations.

By keeping in mind the internal model of GIT, it becomes easier to understand the usual commands and their quirks. By undersanding the design philosophy behind the implementation, the day-to-day usage can become, hopefully, less surprising.