As before, I want you to go into a repository of your choice and poke around a bit yourself while you’re reading my explanations.

The .git directory

In your working directory, run cd .git to visit the .git directory. Here, Git stores everything it needs to run: configuration, the database, hooks and refs. I want to explain every subdirectory briefly, before going into the details of the more interesting parts.

The configuration file config

This file stores options you have configured directly via git config or automatically through other commands. For example, git clone populates this file with the default “origin” remote:

You can edit this file in any text editor. This is sometimes easier than using git config.

The uppercase files

You will notice some files in the .git directory that are named in uppercase letters. Let’s keep this brief, I will get into more detail later.

description, hooks and info

The description file contains a description of your repository. You’ll likely never use this unless you plan to publish your repository through gitweb.

The hooks directory contains callback scripts that are executed by git everytime certain event occurs (like a commit or a rebase). These can be used to send out emails everytime someone pushes a commit to a server for example.

Inside the info directory, the only file you’ll probably ever touch is the excludes file, which contains your private excludes. You can use it to prevent temporary files from showing up in git st without adding them to .gitignore.

index and logs

The index is a central mechanism of Git. Basically it contains the content of your next commit. I like to call the index an unborn commit. By adding and removing files through git add and git rm you shape it’s content to your liking and then store it in the database as a proper commit through git commit.

The logs directory contains specials files known as reflogs. Each of the files here corresponds to a branch. Whenever you are working in that branch, an entry is created in the reflog. This makes it possible to see what commit your branch was pointing to, at any given moment in time using git reflog <branchname>. I will talk a bit more about the reflog in the next part of the series.

The objects directory

Now it gets interesting. The objects directory contains the actual database of all the objects in the repository. The objects are stored in files and directories that are based on the objects SHA1 ids. The first two characters of the SHA1 form a directory, the rest is the filename. If you cat any of the files in there, you’ll see the binary contents of the object, compressed with zlib. To see the uncompressed content, use the following command (you’ll obviously need Ruby for this):

Remember what I told you at the end of part one? That git stores all of its objects as actual files? Here you see them. Also, remember that I told you that it didn’t actually do that all of the time? Well, run git gc and list the contents of the objects directory again. Most of the directories should be gone now. They went into one of the files in objects/pack. These are compressed archives that allow for much more efficient storage of the objects. But it helps to still think of them as the actual files we’ve seen before.

The refs directory

The refs directory sits at the interface between the user and the object database. Here, branches and tags are stored, enabling you to access commits by an easy to remember name instead of the SHA1. Inside refs you’ll see several subirectories: heads and remotes store branches for the local and remote repositories respectively, tags contains tags. If you’ve used git bisect or git stash before, you’ll also find corresponding files for them here.

You can take a look at what your refs are pointing to by just looking at their content. They simply store the hash of the object they’re referencing in plain text.

You might be wondering, that git branch -av is showing you quite a lot more branches than you see files in the refs directory. That’s because only branches you’re actually working with are listed here as files. The rest can be found in the file packed-refs in your .git directory.

Working with branches

Now that you know how branches are stored, you can probably imagine how some of Gits common operations are implemented. Lets take a simple commit for example.

• Let’s assume you’re working in the master branch. Your HEAD will point to that branch. Execute a cat .git/HEAD and you’ll see a reference to the master branch:
  
  ref: refs/heads/master
   
  Master itself might point to a commit:
  
  jan@mops$ cat .git/refs/heads/master
  05c80116a36bbbdd7a453255aee5a1d2c7b01fd7
  jan@mops$ git rev-parse master
  05c80116a36bbbdd7a453255aee5a1d2c7b01fd7
   
  HEAD can either point to a branch, like shown, or directly to a commit (That’s called a detached HEAD, a term you might have encountered already). Git has no problems resolving HEAD to a commit in any case:
  
  jan@mops$ git rev-parse HEAD
  05c80116a36bbbdd7a453255aee5a1d2c7b01fd7
   
  This situation is displayed in the illustration.
• Before you start editing, your working tree, your index and the tree object that belongs to the current commit that HEAD points to have identical content. This is situation 1 in the illustration.
• You will now edit a file. The git status command will report that there’s a difference between your working directory and the index and list the file under “Changed but not updated”. This is situation 2 in the illustration.
• After adding our changes to the index with git add, git st will now report difference between the index an the HEAD under “Changes to be committed”. We’re now at situation 3.
• If you’re done with your work, you finally call git commit. Git then takes your index and creates a tree object from it. A commit object is created, containing the commit message, your name and the current time. The commits parent will be set to the commit that is referenced by the current HEAD and its tree reference will point to the tree that was just created. This is the transition from situation 4 to situation 5.
• Finally, to treat that newly created commit as the new tip of your development history, git updates HEAD to point to it. In case HEAD references a branch, the branch is updated. At every step you can see the pointers changing by looking into your HEAD and refs/* files. You’re now at situation 6 and your repository is in a clean state again.

By now, you can probably already imagine how branches are created. Git simply places a file with the name of the branch in refs/heads and lets it point to the commit you provided to git branch.

Checkouts are a little more interesting. If you instruct git to checkout a branch, three things happen:

• The index is set to the same contents as the commit you’re checking out
• The working directory is also adjusted to the same contents
• If you’re checking out an actual branch (as opposed to, say a tag or a SHA1-identified commit), git updates HEAD to point to that branch.

Now you know what the HEAD file I introduced in the “uppercase files” section is used for. Just as HEAD stores the pointer to your current branch, the other uppercase files point to other branches, or other commits that are interesting in some situations like a merge or fetch operation.

Summary

The last part of the series described the data structures behind Gits object database. By discussing the contents of the .git directory, you understand the operations that git performs to organize the content in the object database, and to create branches. Given the knowledge about these files, you should have a clear idea now how Git implements its commands.

In the next part of the series, I want to take a closer look at some of them, especially the dreaded rebase.

If you used Git before and kinda like it but feel unsure about using some of its advanced commands because you think you don’t completely understand whats going on under Gits hood, if you like what rebase can do for you but are afraid to use it because you’ve read somewhere that the sky will fall on your head if you make a mistake, then this article is for you. Git only reveals its true, awesome power if you use it to its fullest potential. And to do that it is essential to understand how Git works internally.

The talk I gave at the Barcamp was roughly about four topics:

1. Data structures in Git
2. The layout of the .git directory
3. The benefits of using rebase and why rebase isn’t nearly as harmful as everyone thinks
4. Several Tips and Tricks for making day-to-day tasks easier

After the talk I decided to write it down here for the benefit of everyone who ever struggled to understand what Git exactly does when you tell it to pull, merge or commit. I will write a series of four posts, based on the topics of the talk.

Data Structures in Git

Git’s core database is a directed object graph with four different types of objects. Each object has an identifier that is calculated as a SHA1 hash of its contents. That hash is formed by a cryptographic function returning a 128bit key. The fundamental property of a hash function in this context is that it’s a true function in the mathematical sense. Same inputs yield the same outputs. This ensures that identical objects are always assigned to the same identifier. There is no duplication, ever. I will describe this priniciple in more detail in the following paragraphs.

It will help your understanding to refer back to the following graphic when reading these paragraphs: This is a representation of Git’s object graph. For brevity I focused on the structural properties of each object, omitting the content (the binary content in case of a blob, the commit message in commits etc.). I also shortened the SHA1s to three characters.

The four object types in Git are:

Blobs

Blobs are simply chunks of binary data with no other properties, no metadata no nothing. Just the pure data. They are used to store the content of files in the repository. They do not correspond 1:1 to files however. They correspond to file content. Two files in your repository, with different names or at different locations, with identical file content will use the same blob object to represent that content in the database. The blob is identified by the SHA1 hash of its content.

Trees

Trees represent directory structures. This is a tree object:

Where does this representation come from? Well, Git offers a command for that, git cat-file. Its most common usage is git cat-file -p <object>. I got the above printout by passing the SHA1 of a tree as <object>.

You can see that a tree is simply a list of your directory that consists of links to other objects, blobs (for files) and trees (for subdirectories), together with metadata (file permissions and filenames). What this says, for example, is that there’s a file named “Capfile”, whose content is stored in the blob with the SHA1 e04728e8d391f57a6fa0c3325118750c602ef5ef:

Or that “app” is a subdirectory whose content is found in the tree 875c4668c815306dcb1de23407973e2f1fb9d3a8:

Commits

Until now, our trees and blobs have been floating around in the database with no way of getting at any object without knowing its SHA1. Also, we’ve seen the information that blobs and trees can store but there was no discernible way of storing the history of anything. Pretty useless for a version control system, you say?

This is were commits come into play. Their job is to record history. Let’s start by looking at the commit on top of our master branch with git cat-file -p master:

Here you can see what kind of information is stored in a commit:

• An author and a time of authoring as well as a committer and the date the commit was created. This distinction is made because Git supports patches that are authored by one person but committed by someone else, something that’s not uncommon in big open source projects.
  This can also occur when all developers have commit rights: Whenever you cherry-pick a commit, you become the committer, but the original author remains the same. However, for the sake of discussing Git’s data structure this distinction has no relevance.
• A reference to a tree object that represents the state of the working directory at the time the commit was created.
• One or more references to parent commits. This is what actually builds the history of your repository. Each regular commit has one parent, one previous state of the working directory. When you perform a merge, a commit can even have two or more parents, pointing to the different branches of development that have been merged.

Taking the example from above, if we inspect the parent we see such a merge commit:

This explains how Git strings a series of commits together to form a history, but I didn’t tell how to actually get at an object without knowing its SHA1. This is were branches come into play. They are essentially readable aliases for SHA1s that get updated every time you perform certain actions (like committing, merging, etc.). I will explain this in more detail in the next post of this series.

Tags

The last type of objects are tags. To be more specific, Annotated Tags. Simple tags are not objects (more on that later), but annotated tags are. You get an annotated tag if you use the -a option when creating a tag

A tag consists of an object reference, but what’s special about it, is that it can refer to any kind of object. What exactly the tag is referring to, can be seen in the type field. The tag we’re seeing here points to a commit with the SHA1 4a06c46ee6d58ce4be09954ee054921b18269cd6. The tag also has a name, given in the tag field, a tagger and a message.

To be honest I never worked with annotated tags and most of you probably never will either. Their main use case over regular tags is that they can be cryptographically signed (as can commits).

Summary

That was it. Four very simple types of objects.

• Blobs – Storing the content of files
• Trees – Storing the structure of your working directory
• Commits – Putting trees into a sequence to preserve history
• Tags – Reliable mechnism to point to objects in the database

Maybe you should take a look at one of your own repositories now, starting with git cat-file -p master and poking around a bit.

These objects and their references to each other are the absolute core of Git and understanding their structure and relationships is essential to working well with Git. As soon as you start thinking of your repository as this objectgraph, you’ll realize that the Git toolchain is nothing but a set of manipulations on that graph database, creating new objects all the time, pointing to other objects.

There are two additional implementation details that you should be aware of:

1. Even though my description and the results of git cat-file make it seem like you’re dealing with full fledged objects, the reality is that Git uses very efficient compression algorithms to reduce the amount of actual data stored in its database. Trees or blobs aren’t usually stored in full but described as differences to other similar objects.
   But for reasoning about objects, you can and should think of them as being self-contained and independent.
2. On the other hand, objects often actually are uncompressed in the database. Also Git never deletes objects! If you remove a file from a tree, the blob for that file will still exist. If you lose a commit through rebasing or merge problems or by accidentally deleting a branch, as long as you know its SHA1, you can get back to it. The only time Git actually deletes and compresses objects is during its garbage collection run. If you clone a new project, you will retrieve compressed objects from the remote server, but objects you create in your local repository will be uncompressed at first. Git repacks and garbage collects on its own from time to time, but you can also trigger this processs manually by calling git gc. Do this when you notice that working with your repository becomes slow or that the repository becomes too large.

Next post

In the next post of this series, I will explain the structure of the .git directory. This is where you will find your branches and regular tags, as well as the actual object database files. You will learn what a branch actually is and how to effectively manipulate them.

Download/View Slides

The talk has also been recorded by Oliver Überholz, who promised to send me a copy but so far hasn’t replied to my messages. Please give him a nudge :)

UPDATE: First post of of the Advanced Git series is online.

I’ve updated the README and the project page with the new information.

I’ve also published my dbserialize plugin at Github.

Git

Shortly before Git became really popular some months ago, I had become interested in darcs and distributed revision control systems in general. The topic is kinda difficult though and none of the texts I was reading at the time could really communicate the benefits of DRCS to me. I always had some gripes about svn but it wasn’t clear to me how DRCS were able to solve them.

I lost interest, following posts about git only loosely until last night a colleague pointed me to Randal Schwartz’ Git presentation at Google Tech Talks. Holy crap, I need to check this out. What appeals most to me:

• The ability to have the entire repository available locally
  I was extremely sceptical when I first heard about this, but when Schwartz claimed that the entire repository of the linux kernel is half the size of a checkout I was sold.
• Subversion interoperability
  Didn’t know about this before. Makes the transition much easier.
• Having local-only repositories inside your working dir
  I have many smaller projects that I’d love to keep locally contained. In Subversion I always had to create a repository on my server for everything.
• Other small things
  High compression, the simple database system behind git, the optimization for speed, staged commits, the ability to completely erase files from a repository (e.g. stuff not intended for publication, something that was very hard to do on subversion), the placement of all metadata in a single directory (opposed to littering every dir in the working copy with .svn directories)

Despite some shortcomings that was enough to make me install git on my mac (sudo port install git-core). I’m eager to check it out later today.

Smalltalk and Seaside

Ever since watching Evan Phoenix Rubinius Presentation at RubyConf 2007 and listening to Avi Bryants Smalltalk’s Lessons for Ruby Keynote from RailsConf 2007 I’ve been curious about Smalltalk. I mean, I was curious about it before, after all it’s probably the language that has the most influence on what I’m doing today (through its promotion of object-orientation and through providing key principles behind ruby), but since listening to Avi and Evan I’ve become really interested in VM implementations (see Smalltalk-80: The Language and Its Implementation for an excellent in-depth description of the orignal Smalltalk-80 interpreter) and real world usage of smalltalk.

To be honest, as much as I love Ruby as a language, its implementations all suck. And Evan explained why: Implementing most of the base language on another Platform (C for MRI, Java for JRuby) turns out to be a leaky abstraction when you want to extend the language. Additionally, as pure and beautiful the Ruby language is in concept, as ugly is its implementation. On the one hand, what I like so much about Ruby is its conceptual purity, its very limited set of axioms, syntax and exceptions from its own rules, on the other hand, this purity is not present in the interpreter when high-level data structures (like arrays and hashes) are implemented in C for performance reasons. Smalltalk has always had a strong philosophy of implementing as much as possible in Smalltalk itself and only resorting to C for a minimal subset (“Turtles all the way down”).

Rubinius aims to implement a Ruby interpreter on the design principles of smalltalk. I love the project and there seem to be only the most brilliant people working on it (Evan Phoenix, Eric Hodel, Ryan Davis and others, full time). As many others have said already, Rubinius is likely to become the main Ruby implementation if they manage to take off (and they will undoubtedly).

Yet, something was bugging me: If Ruby and Smalltalk are so similar, if Smalltalk has been around, specified and stable for 25 years in many different, compatible implementations, commercial and open source, why take the long route and bend Ruby to look like Smalltalk? Why not use Smalltalk directly? These questions became even more nagging after reading Randal Schwartz’ (yeah, the guy who sold me on Git earlier) Transcript show: ‘Hello, world!’, cr and Fabio Akitas excellent Interview with Avi Bryant (part 1, part 2). Listening to yet another chat with Avi (linked in the second part of Fabio’s interview) at Floss Weekly (with Randal Schwartz again, highly recommended) before falling asleep I finally decided to check out Squeak, the (most popular?) opensource Smalltalk implementation. I was amazed at the simplicity of the installation: Download the VM, Download the Squeak image, load the image into the VM, done.

I have read about Seaside, Avi’s web development framework, before, mainly in reagard to it’s clever use of continuations, but some of the stuff he described at Floss sound almost too good to be true. Live debugging of your app in the browser ? With hot code swapping over the wire? And I thought Rails’ rdebug integration was great.

Well, at 2am I was finally falling asleep but the stuff I’ve been discovering will probably keep me occupied for quite some time. As I explore and discover more about the topics mentioned, I’ll report my findings here on my blog.