Made of Bugs

On Git and Usability

I've been helping a number of people get started working with git over the last couple of weeks, as Ksplice has brought on some new interns, and we've had to get them up to speed on our internal git repositories. (As you might expect from a bunch of kernel hackers, we use git for absolutely everything). While that experience is what prompted this post, it wasn't really anything I haven't seen before as SIPB transitioned from a group that mostly versioned code in SVN or SVK to one that used git almost exclusively, practically overnight, as these things go.

I love git, and use it everywhere. One of the things I particularly love about git is the fact that I understand essentially every element of git's data model and how it works, which means that I can invent odd workflows or script git to do things that I never could have with SVN or even SVK. Part of the reason for this is that git, internally, really is fundamentally simple in a way subversion isn't. There's some complexity (which I don't and don't have to understand) behind implementing its model efficiently, but I can describe the basic git data model in about a single slide (And in fact I have).

At the same time, though, I consistently find that people who come to git from subversion find it insufferably complex and hard to learn, and spend quite a while wrangling with it, before they eventually fight it to a draw where they only need to ask their local git expert to dig them out of a hole every once in a while, or else until they just give up and declare git to be space-alien and go back to something they "understand".

The problem

Having helped enough of these people, I've begun to understand the problem. Fundamentally, the way (most) people learn and think about subversion is different from the way git experts think about git. subversion's internal model is fairly complex, but you are not expected to understand it. You just have to know the half dozen commands you'll ever need, and you use them, and everything is fine.

Git's model is fundamentally fairly simple (a DAG of immutable commit objects where branches are named mutable pointers into it), but you are expected to understand it fully to use git effectively.

A subversion user who wants to get the latest updates from the server just knows to run svn up and go back to their life. A git user has to (in general) think about whether they want to rebase their local changes onto the remote changes or merge them, and which remote branch they want (this is all, of course, expressed essentially in terms of the git data model), and then maps that back into the commands they want to perform this task.

Subversion users aren't used to thinking like this, and so when they ask a git user for help, they get one of two classes of answers, neither of which they like. Either they just get back a list of half a dozen possible commands they could choose (since the git user has mapped their request into the git object model, and can think of a bunch of ways to implement that operation, depending on your mood and the phase of the moon), or else they get a "What are you really trying to?", since their request has multiple interpretations in the language of git objects. Neither answer is a single command that will always solve their problem, which is what they want, so they can just go back to whatever they were trying to do, and so they end concluding that git is needlessly strange and confusing.

I think there are at least two reasons for this different perspective from git users. One is historical – git was designed to work this way. Linus wrote git as a dumb content tracker, and intended other people to write various UIs on top of it, but not really for anyone to use it as-is. So fundamentally, git is designed to just implement this basic model, and not to export a single way to do anything. The other, perhaps more fundamental reason, is that git is designed to be infinitely flexible, and so it's fairly rare that you can give an answer for "How do I do X with git?", since the answer will often depend on, "Well, what are your project's conventions?"

What do we do about this?

So, as git users, hackers, and evangelists, what can we do to improve this situation? I think that basically, we need to understand what's going on here, and try to have better answers for what the subversion users regard as simple questions. We should encourage people to get to know and love the git object model, but realize that most of the time, they just want to get work done, and we need to try to have simple answers to simple questions.

To that end, I think there are three main remaining areas in the git UI that I think new users find confusing, and that we need to figure out how to improve. These are pulling and pushing changes, and the whole issue of the git index. Here are my thoughts on what's wrong, and what we can improve on:

  • ''git-push'': This command is a disaster in a number of ways. matching is utterly the wrong default for push.default (I believe 1.7 is going to fix this). The git push $REMOTE $BRANCH syntax is not optimized for the common "subversion-like" case of a single remote, and leads countless users to attempt to git push master, at which point they get the less than totally helpful error message:

    fatal: 'master' does not appear to be a git repository
    fatal: The remote end hung up unexpectedly
    

In general, the error messages are pretty bad. I believe it was @defunkt who nominated "src refspec does not match any" as the "worst error message of all time", and it's hard not to see his point.

  • ''git-pull'' is also problematic. The vast majority of users in a project live at the edges, pushing code inwards, and they almost never actually want to create a merge commit. I almost always end up telling users "No, really, you always just want pull --rebase", which leaves them with a poor impression of git's UI (if you always want that flag, why isn't it just mandatory?), and which is hard to explain, because the whole concept of "rebasing" is difficult to really grok without really understanding the commit DAG.

branch.<name>.rebase and branch.autosetuprebase are a partial solution, but is there a reason I can't just have a repository wide default option that makes --rebase the default? With the former solution, I'm never confident I've actually set it on all of my branches, or that it will get set if some script creates a branch without using git-branch or git-checkout.

Similarly, it sucks that you can't pull if you have any uncommitted changes at all, especially for subversion users who are used to this Just Working.

  • Finally, we really need to improve the story with the index. As an experienced git user, I love the index, and use it for many reasons all the time. However, as anyone who has tried to explain it (or even to learn git) probably knows, it's a confusing idea at first.

Generally, when someone is starting out with git, I try to ignore the index, and suggest they just always use git commit with a path. I then have to handwave, of course, over why they need git commit -a instead of git commit, but that's a mostly minor problem.

The bigger issue comes when a user finds themselves in a merge conflict, and finds that it's basically impossible to do anything without fully understanding the index, and all of the weird commands for moving content between the index, the working tree, and HEAD. Just the other day, three of us git experts had to stop and think for a while about how to answer the question "So, if I'm in a conflicted merge, and I want to just conditionally take “their” changes, what the hell do I do?". That's not a good sign for new users being able to figure out what to do here.

I don't fully know what the right answer is, but I think it's clear the tools need some help here.