Explaining why git is evil
Latest commit 618251e Apr 6, 2013 @regebro Update README.md
Failed to load latest commit information.
README.md Update README.md Apr 6, 2013


Please note that github is not a bloody discussion forum. 'K?


OK, so I've created a new repo on github, and I've cloned it:

$ git clone  git@github.com:regebro/die-git-die.git

I add my code to the repo and commit and push:

$ git commit -a -m"Commit."
$ git push

Now I want to add some github pages for this repo. This is done by the weird practice of creating a special branch called gh-pages and then deleting everything on that branch.

Yes, really. Kinda scary. But it shouldn't be a problem, should it? Because we are using a code revision system, so even if something gets messed up, it's always possible to go back to an earlier revision and do it from there.

$ git clone git@github.com:regebro/die-git-die.git die-git-die.pages
$ cd die-git-die.pages
$ git checkout --orphan gh-pages
$ git rm -rf .

So we create some pages to have there. It's HTML.

$ echo "<html><body><h1>Die git, die\!</h1></body></html>" > index.html
$ git add .
$ git commit -a -m "First pages commit"
$ git push origin gh-pages

While doing up these pages, we find some errors in the original branch and fix them, which we push and commit.

$ cd ../die-git-die
$ git commit -a -m"Commit"
$ git push 

And let's then clean up the HTML in the pages a bit:

$ cd ../die-git-die
$ vi index.html
$ git commit -a -m"Commit"
$ git push 

! [rejected]        master -> master (non-fast-forward)
error: failed to push some refs to 'git@github.com:regebro/die-git-die.git'
To prevent you from losing history, non-fast-forward updates were rejected
Merge the remote changes (e.g. 'git pull') before pushing again.  See the
'Note about fast-forwards' section of 'git push --help' for details.

Oi! What? Why? There's no conflicts, there can't be any conflicts, the two branches doesn't even have one single file in common. Why didn't it want to push? But, eh, OK, I'll pull.

$ git pull

You asked me to pull without telling me which branch you
want to merge with, and 'branch.gh-pages.merge' in
your configuration file does not tell me, either. Please
specify which branch you want to use on the command line and
try again (e.g. 'git pull <repository> <refspec>').

What!? Doesn't it know I'm on the gh-pages branch?

$ git checkout gh-pages
Already on 'gh-pages'

Yes, it does. So why does it ask me with which branch I want to merge!? That makes no sense. But OK, I'll tell it:

$ git pull gh-pages
fatal: 'gh-pages' does not appear to be a git repository
fatal: The remote end hung up unexpectedly

Aha, OK.

$ git pull git@github.com:regebro/die-git-die.git gh-pages
From github.com:regebro/die-git-die
 * branch            gh-pages   -> FETCH_HEAD
Already up-to-date.

Huh. This is extra strange, since on another repo I have, it does know what to pull from automatically. The error message tells me to fix my config file, and I notice that the other repos config file does include those changes, so I'll do that here as well. No idea why it worked automatically once.

[branch "gh-pages"]
    remote = origin
    merge = refs/heads/gh-pages
    rebase = true

OK, now git pull works. But it didn't merge anything, so... I'll try to push again:

$ git push 

! [rejected]        master -> master (non-fast-forward)
error: failed to push some refs to 'git@github.com:regebro/die-git-die.git'
To prevent you from losing history, non-fast-forward updates were rejected
Merge the remote changes (e.g. 'git pull') before pushing again.  See the
'Note about fast-forwards' section of 'git push --help' for details.

No? OK, so I'll force it then. It tells me I can lose history, but this is a revision control system, surely they don't mean I can actually lose history?

$ git push -f
Counting objects: 6, done.
Delta compression using up to 2 threads.
Compressing objects: 100% (4/4), done.
Writing objects: 100% (6/6), 626 bytes, done.
Total 6 (delta 0), reused 6 (delta 0)

OOPS! Everything you did on the master branch since you made the gh-pages branch now dissapeared! Gone! Poof!

You can get it back by going over to the other clone and pushing it again, because it still exists locally. But that's not obvious. But don't do a pull there first!

$ cd ../die-git-die
$ git push

And if you pull in that clone first, well, then you are in deeper trouble, but you can still get it back because it's still exists in the local history, although I'm unsure how.

If you delete that local clone though, it's gone forever.

So what is the reason for this? What did I do wrong? Well, the problem seems to be that git by default will push "matching" branches, ie all branches that exist both locally and remotely. However, git will not pull matching branches. That means that all your branches except your current branch will not get updated when you pull. And when you push, your local git will then, in a fit of amazing stupidity, try to push your current state of these other branches onto your origin.

This is so mindboggingly stupidly inconsistent that it blows my minds. Still. Two weeks after I learned this.

Git 2.0 will change the default to current, it seems, which solves the problem at least partly. Meanwhile you can add the following addition in your .gitconfig to set a more sane default.

    default = current

I can however not find any way to set the same thing for pull, so what it pulls for default is still somewhat unclear to me.

Of course, a less good but at least sane behavior would have been if pull also pulled matching branches by default. But it doesn't. And it seem like there is no way to make it do so either, which also boggles the mind. Pull has an "--all" parameter, which seems to make no discernable difference.

It is in any case completely obvious that pull and push should pull and push the same branches. All other behavior is just confusing and will cause problems. Currently git does not do this by default, and I'm not even sure it's possible to configure it so it does.