This is part one of a series:

  1. Unlimited growth (in a bad way)
  2. How we stopped making everything worse
  3. Dealing with distribution

In order to understand this story, we'll need an appropriate mental model of Git. At their core, version control systems save previous states of a project. That way programmers can wind back time to figure out how bugs were introduced. An analogy could be recording a chess game so that interested observers could analyze important moves. Take, for instance, Deep Blue versus Kasparov, 1997, Game 6.

The traditional approach is to record each move:1

1.e4 c6 2.d4 d5 3.Nc3 dxe4 4.Nxe4 Nd7 5.Ng5 Ngf6 6.Bd3 e6 7.N1f3 h6 8.Nxe6 Qe7 9.O-O fxe6 10.Bg6+ Kd8 11.Bf4 b5 12.a4 Bb7 13.Re1 Nd5 14.Bg3 Kc8 15.axb5 cxb5 16.Qd3 Bc6 17.Bf5 exf5 18.Rxe7 Bxe7 19.c4 1-0

Finding the board state at a moment in time requires beginning at a known state (usually the standard setup) and applying one move at a time. This is pretty much the approach used by traditional source control systems such as SCCS, RCS and their many decedents. It's called the changeset storage method and it's the most common approach used by version control systems. The system saves the difference between individual files each time they are changed.

Git, by contrast, saves a snapshot of the entire project at each state in the past. In chess, the equivalent would be saving a copy of the complete board after every move:2

The first four moves of Deep Blue versus Kasparov, 1997, Game 6 displayed as 5 separate boards.

There's even a way to encode board states with the delightful name of Forsyth Notation. Here are the first four moves of game 6 using that system:3


Obviously this approach requires much more storage space. A lot of that space is used to record pieces (such as the a2 and a7 pawns) that are rarely moved. The advantage is that it's a lot easier to return to a previous state. Instead of applying changesets one at a time, you can get previous states all at once.

Now Git has some tricks up its sleeves to reduce the storage problem. According to the Pro Git:

Git is a content-addressable filesystem. Great. What does that mean? It means that at the core of Git is a simple key-value data store. What this means is that you can insert any kind of content into a Git repository, for which Git will hand you back a unique key you can use later to retrieve that content.

To go back to the chess analogy, suppose I wanted to store many games where each move is a snapshot of the board. One way to do this would be to assign a number to each unique board position as they are added. So the initial state could be labeled 0 and the state after 1. e4 could be called 1. After storing the first four moves of game 6, I'd have a table that looked like this:

Key Value
0 rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR
1 rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR
2 rnbqkbnr/pp1ppppp/2p5/8/4P3/8/PPPP1PPP/RNBQKBNR
3 rnbqkbnr/pp1ppppp/2p5/8/3PP3/8/PPP2PPP/RNBQKBNR
4 rnbqkbnr/pp2pppp/2p5/3p4/3PP3/8/PPP2PPP/RNBQKBNR

This is convenient because every game starts with position #0 instead of the 44 character Forsyth string. Then if white uses the most common opening move, the next position is #1. Now for chess, this strategy is limited because there are something like 2.2 * 1046 board diagrams and the maximum Forsyth code is 71 characters. So this sort of coding wouldn't save all that much space.

Fortunately, the situation is different for Git repositories. Every file, no matter how large, can be assigned a 40-digit hexadecimal number that is (quite likely) to be unique to that file.4 The keys are generated using the Secure Hash Algorithm 1 designed for cryptography. The number (usually called a "SHA" and pronounced "Shaw") is generated from each file's contents and even the smallest change results in a new number. So when a new file is added, Git can generate a key and look up in the index whether another file with the same contents has already been stored. And that means every version of every file, no matter how large, is stored only once.5

Git uses SHAs everywhere. When you commit changes, Git takes a snapshot of the working directory and generates a unique SHA for that commit. Again, this number is (virtually) guaranteed to be unique to the exact details of the state of the working directory. This is important because it means any change to the working tree will result in a different commit. For more details, see "Commits are snapshots, not diffs" on the GitHub blog.

Failing to heed the warnings

Pro Git helpfully explains:

There are a lot of great things about Git, but one feature that can cause issues is the fact that a git clone downloads the entire history of the project, including every version of every file. This is fine if the whole thing is source code, because Git is highly optimized to compress that data efficiently. However, if someone at any point in the history of your project added a single huge file, every clone for all time will be forced to download that large file, even if it was removed from the project in the very next commit. Because it’s reachable from the history, it will always be there.

Truth is, Git does just fine storing large files. I use it for images on this very blog and I saved episodes of a podcast I did at Stack Overflow directly in my repository. It's handy to have these binary files stored along side of the rest of those projects. Yes they make the repository larger, but you gotta store them somewhere.

At my job as a Developer Advocate for EDB, I've been looking at speeding up the build of our new documentation site. It can take over an hour with 6 minutes or so of that was just checking out the branch from the repository. Depending on what was cached, it might take a little more or a little less, but that's a lot more time than it really should take.

Cloning the entire repository took a lot of time too. If your internet connection isn't especially fast, it could take 45 minutes to clone. The reason isn't complicated. Our repository weighed in at 10 gigabytes. A large portion of that sum came from screenshots, icons and other miscellaneous images. Yes, they were larger than the source files Git is designed to store, but they were not the bulk of the problem. The real problems started when we committed PDF versions of our documentation.

Git isn't optimized to deal with images, but they tend to be static. Every once in a while you might redo a screenshot or a diagram, so it's a bit of overkill to keep them in Git, but not the worst. On the other hand, we generate PDFs from our HTML documentation. Every time we merged a branch into develop, we built all the PDFs at the same time so they would match the web versions. And then we committed the changes to Git.

For those of you who haven't turned away from the screen in revulsion, let's review why this is, to use polite words, sub-optimal. Every time we committed a change to our documentation, Git would compare all the staged files in our working directory to the copies that already exist in the repository. It would do this using a cryptographic key (SHA-1) that registered even the smallest change as a completely new file. Then it would save the entire project as a single commit. At any point in time we could go back to each version of the PDFs, so Git keeps every copy forever. And it's not very good at compressing this type of file.

This is unlimited linear growth. Each time we built out we increased the size of the repo by roughly the size of our PDFs after zlib compression. By design, Git repositories never get smaller. It's a ratcheting effect that needless chewed up disk space and bandwidth for everyone who contributes to the project. It had to stop.

Next time: we stop digging.


  1. The full Portable Game Notation (PGN) for this game is:

    [Event "IBM Man-Machine"]
    [Site "New York, NY USA"]
    [Date "1997.05.11"]
    [EventDate "?"]
    [Round "6"]
    [Result "1-0"]
    [White "Deep Blue (Computer)"]
    [Black "Garry Kasparov"]
    [ECO "B17"]
    [WhiteElo "?"]
    [BlackElo "?"]
    [PlyCount "37"]
    1.e4 c6 2.d4 d5 3.Nc3 dxe4 4.Nxe4 Nd7 5.Ng5 Ngf6 6.Bd3 e6 7.N1f3 h6 8.Nxe6 Qe7 9.O-O fxe6 10.Bg6+ Kd8 11.Bf4 b5 12.a4 Bb7 13.Re1 Nd5 14.Bg3 Kc8 15.axb5 cxb5 16.Qd3 Bc6 17.Bf5 exf5 18.Rxe7 Bxe7 19.c4 1-0

    Most of this code is metadata with only the final line (1.e4 c6 2.d4 d5 ...) being the actual recording of the game. Version control systems carry around a lot of meta data too, but it's not important to our story just now.

  2. This is a screenshot from 5D Chess With Multiverse Time Travel which is a not-entirely successful chess varient. Since it shows a certain type of branching, I suspect I'll use it in future posts of this series.

  3. Again, I've simplified this from the full Forsyth-Edwards Notation (FEN):

    rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
    rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq e3 0 1
    rnbqkbnr/pp1ppppp/2p5/8/4P3/8/PPPP1PPP/RNBQKBNR w KQkq - 0 2
    rnbqkbnr/pp1ppppp/2p5/8/3PP3/8/PPP2PPP/RNBQKBNR b KQkq d3 0 2
    rnbqkbnr/pp2pppp/2p5/3p4/3PP3/8/PPP2PPP/RNBQKBNR w KQkq d6 0 3

    The first part of each line (up to the space) was originally invented by David Forsyth in 1883 to facilitate his chess column in the Glasgow Weekly Herald. Stephen J. Edwards expanded the system to include:

    • Which side is currently on the hook to move.
    • Whether each king can castle and on which side(s).
    • If a pawn is open to en passant capture.
    • The number of moves since the last capture or pawn advance.
    • The curent move number.

    All of this meta data is necessary for giving chess programs the complete state of the game at that moment in time. In the Git analogy, the full FEN notation is something like a commit and the simple board diagram is like the current working tree.

  4. Technically Git saves the contents of a file as a blob. The distinction is that a blob doesn't include the filename. That means you can change the name of a file in Git without generating a new copy of the contents.

  5. Git also compresses blobs and uses packfiles to reduce the size of the repository. Interestingly, packfiles work on a similarly to changesets except they can group similar content even if it's not strictly speaking the same file.