Hacker News Reader

Comment by gcv | Hacker News Reader

Commits happen Subversion-style. Yes, this is a feature! No need to add files before every commit or use a flag.

Please, Git maintainers, don't listen to this guy and don't ever change the current behavior. I've come to rely on staging my commits incrementally.

I really care about keeping a pristine history with isolated and understandable commits. Being able to commit pieces without including temporary code or debugging comments is invaluable to my workflow. I want to be able to modify source code without worrying about any piece of it getting committed accidentally. I also want all edits I made, including uncommitted ones, to stick around until I explicitly remove them.

Maybe Mercurial can do this also, but Git has been almost completely do-what-I-mean with regard to making sure I don't lose any changes or confuse myself about where they went.

View on HN

With Mercurial, you can either use queues and the "qref" command to continually prep a commit in progress, or use the "record" extension (like an interactive Git commit) to pick and choose parts of the working tree to commit.

I suppose it's a matter of what the common case is. For me, the majority of my commits are either one or two files, or the whole working tree -- so Mercurial's defaults make sense with that usage pattern. Of course, other usage patterns would dictate more of a "massage the commit" style.

Git's incremental commits may be a great idea in theory, but in practice, it allows you--even forces you, if you run in any mode except "git add ."--to make commits you've never actually tested. I'd rather have a cluttered history that built at any given point than a clean one that didn't.

I feel the opposite way. I often see hunks like:

     sub foo {
    +    warn "made it here";
         actual code;
     }

This stuff, I never want to commit. Sure, I haven't tested the code without it, but I'm sure it works.

Subversion-style committing sucks hard; worst idea ever. (And besides, svn still lets you do this; svn commit one-file. Untested tree committed, with no possibility of ever backing out the change. Not good.)

It's not so bad as long as you don't push any untested changes. If I do something freaky with the index, I usually stash the leftovers and test the last commit. Like this:

  git add -i
  git commit
  git stash
  ./some_tests  # ...shoot, forgot something
  git add .     # whatever I forgot
  git commit --amend
  ./some_tests  # now we're good
  git stash apply
  git stash clear
  git push

On the hg side, you can use mq and record to get yourself in and out of the same kind of trouble, as someone else mentioned. I think these are essential but dangerous features that any good DVCS should have.

A better safety mechanism (than deleting the feature) is a pre-push hook that runs the test suite on every commit since the remote head. Not everyone needs that level of strictness, though.

It's the other way around. Partial commits terrify me in theory; in practice it's really easy to work out which lines you want to commit and which lines you don't, and having a clean history is really really nice.

And since you've got both "git commit --amend" and "git rebase -i" to hand, it's trivial to later on fix up any mistakes you do make.

Honestly, having a clean history is an enormous win - it far outweighs the minor cost of making the odd wrong but easily-correctable-after-the-fact commit.

You can have both, you know: you use git stash/Mercurial attic to shelve the parts of the change you don't want, test the result, commit, then reapply what you stashed. You get a clean history, and you can easily verify your source code works. This solution notably does not involve explicitly exposing the dirstate to the user.

Reply on news.ycombinator.com