Missing Semester Lecture 6 - Version Control (Git)

MIT The Missing semester Lecture of Your CS Education Lecture 6 - Version Control (Git)

Modeling history: relating snapshots

How should a version control system relate snapshots? One simple model would be to have a linear history. A history would be a list of snapshots in time-order. For many reasons, Git doesn’t use a simple model like this.

In Git, a history is a directed acyclic graph (DAG) of snapshots. That may sound like a fancy math word, but don’t be intimidated. All this means is that each snapshot in Git refers to a set of “parents”, the snapshots that preceded it. It’s a set of parents rather than a single parent (as would be the case in a linear history) because a snapshot might descend from multiple parents, for example, due to combining (merging) two parallel branches of development.

Git calls these snapshots “commit”s. Visualizing a commit history might look something like this:

o <-- o <-- o <-- o
            ^
             \
              --- o <-- o

In the ASCII art above, the os correspond to individual commits (snapshots). The arrows point to the parent of each commit (it’s a “comes before” relation, not “comes after”). After the third commit, the history branches into two separate branches. This might correspond to, for example, two separate features being developed in parallel, independently from each other. In the future, these branches may be merged to create a new snapshot that incorporates both of the features, producing a new history that looks like this, with the newly created merge commit shown in bold:

o <-- o <-- o <-- o <---- o
            ^            /
             \          v
              --- o <-- o

Commits in Git are immutable. This doesn’t mean that mistakes can’t be corrected, however; it’s just that “edits” to the commit history are actually creating entirely new commits, and references (see below) are updated to point to the new ones.

---

Data model, as pseudocode

It may be instructive to see Git’s data model written down in pseudocode:

// a file is a bunch of bytes
type blob = array<byte>

// a directory contains named files and directories
type tree = map<string, tree | blob>

// a commit has parents, metadata, and the top-level tree
type commit = struct {
    parents: array<commit>
    author: string
    message: string
    snapshot: tree
}

It’s a clean, simple model of history.

---

Objects and content-addressing

An “object” is a blob, tree, or commit:

type object = blob | tree | commit

In Git data store, all objects are content-addressed by their SHA-1 hash.

objects = map<string, object>

def store(object):
    id = sha1(object)
    objects[id] = object

def load(id):
    return objects[id]

Blobs, trees, and commits are unified in this way: they are all objects. When they reference other objects, they don’t actually contain them in their on-disk representation, but have a reference to them by their hash.

For example, the tree for the example directory structure above (visualized using git cat-file -p 698281bc680d1995c5f4caaf3359721a5a58d48d), looks like this:

100644 blob 4448adbf7ecd394f42ae135bbeed9676e894af85    baz.txt
040000 tree c68d233a33c5c06e0340e4c224f0afca87c8ce87    foo

The tree itself contains pointers to its contents, baz.txt (a blob) and foo (a tree). If we look at the contents addressed by the hash corresponding to baz.txt with git cat-file -p 4448adbf7ecd394f42ae135bbeed9676e894af85, we get the following:

git is wonderful

---

References

Now, all snapshots can be identified by their SHA-1 hashes. That’s inconvenient, because humans aren’t good at remembering strings of 40 hexadecimal characters.

Git’s solution to this problem is human-readable names for SHA-1 hashes, called “references”. References are pointers to commits. Unlike objects, which are immutable, references are mutable (can be updated to point to a new commit). For example, the master reference usually points to the latest commit in the main branch of development.

references = map<string, string>

def update_reference(name, id):
    references[name] = id

def read_reference(name):
    return references[name]

def load_reference(name_or_id):
    if name_or_id in references:
        return load(references[name_or_id])
    else:
        return load(name_or_id)

With this, Git can use human-readable names like “master” to refer to a particular snapshot in the history, instead of a long hexadecimal string.

One detail is that we often want a notion of “where we currently are” in the history, so that when we take a new snapshot, we know what it is relative to (how we set the parents field of the commit). In Git, that “where we currently are” is a special reference called “HEAD”.

---

Repositories

Finally, we can define what (roughly) is a Git repository: it is the data objects and references.

On disk, all Git stores are objects and references: that’s all there is to Git’s data model. All git commands map to some manipulation of the commit DAG by adding objects and adding/updating references.

Whenever you’re typing in any command, think about what manipulation the command is making to the underlying graph data structure. Conversely, if you’re trying to make a particular kind of change to the commit DAG, e.g. “discard uncommitted changes and make the ‘master’ ref point to commit 5d83f9e”, there’s probably a command to do it (e.g. in this case, git checkout master; git reset --hard 5d83f9e).

---

Staging area

This is another concept that’s orthogonal to the data model, but it’s a part of the interface to create commits.

One way you might imagine implementing snapshotting as described above is to have a “create snapshot” command that creates a new snapshot based on the current state of the working directory. Some version control tools work like this, but not Git. We want clean snapshots, and it might not always be ideal to make a snapshot from the current state. For example, imagine a scenario where you’ve implemented two separate features, and you want to create two separate commits, where the first introduces the first feature, and the next introduces the second feature. Or imagine a scenario where you have debugging print statements added all over your code, along with a bugfix; you want to commit the bugfix while discarding all the print statements.

Git accommodates such scenarios by allowing you to specify which modifications should be included in the next snapshot through a mechanism called the “staging area”.

---

Exercises

1. Clone the repository for the class website.
   1. Explore the version history by visualizing it as a graph.
   2. Who was the last person to modify README.md? (Hint: use git log with an argument).
   3. What was the commit message associated with the last modification to the collections: line of _config.yml? (Hint: use git blame and git show).

git clone https://github.com/missing-semester/missing-semester.git

git log --all --graph --decorate

git log -1 --pretty=format:'%an' -- README.md

git blame _config.yml | grep 'collections:' | awk '{ print $1 }' | git show --pretty=format:'%B' --no-patch

2. One common mistake when learning Git is to commit large files that should not be managed by Git or adding sensitive information. Try adding a file to a repository, making some commits and then deleting that file from history (you may want to look at this).

git filter-branch --force --index-filter \
"git rm --cached --ignore-unmatch PATH-TO-YOUR-FILE-WITH-SENSITIVE-DATA" \
--prune-empty --tag-name-filter cat -- --all

4. Clone some repository from GitHub, and modify one of its existing files. What happens when you do git stash? What do you see when running git log --all --oneline? Run git stash pop to undo what you did with git stash. In what scenario might this be useful?

git stash
Saved working directory and index state WIP on master: 159d10a fix typos (#292)

git log --all --oneline
5a59b6c (refs/stash) WIP on master: 159d10a fix typos (#292)

git stash pop
Dropped refs/stash@{0} (5a59b6c9775ebb2c3036d673e21499adb8f973e4)

git-stash would be useful when checking out to other commits but do not want to overwrite the local changes not yet committed, either because they don’t seem good enough yet to commit or there are more urgent bugs to address at the moment, etc.

4. Like many command line tools, Git provides a configuration file (or dotfile) called ~/.gitconfig. Create an alias in ~/.gitconfig so that when you run git graph, you get the output of git log --all --graph --decorate --oneline. You can do this by directly editing the ~/.gitconfig file, or you can use the git config command to add the alias. Information about git aliases can be found here.

git config --global alias.graph 'log --all --graph --decorate --oneline'