git default behavior seems odd from a Unix command line point of view

Hi all -

I am recently working within git after some experience with mercurial.
 I am observing what I believe to be an odd default behavior from the
perspective of UNIX command line tools.  I thought I'd share in case
this hasn't occurred to git maintainers or in case somebody has
developed good workaround practices.

ais@ace:bio[1]$ pwd /home/ais/repo/nps/projects/bio ais@ace:bio$ git status # On branch master nothing to commit (working directory clean) ais@ace:bio[1]$ git commit -a # On branch master nothing to commit (working directory clean) ais@ace:bio[1]$

The [1] in my prompt indicates the exit code of the git commands. What I find odd is that even with the -q option, you get this verbose output. Also, you get a non-zero exit status (which I would expect only on a failure such as presence of an unresolved conflict). My git usage is to have a number of small repositories and use a shell script to loop over them and perform a sync with a centralized server. Having all this wordy output on a "no sync necessary" scenario seems counter the desired properties of output only when work is taking place or when an error occurs.

Have others developed git practices to sync a bunch or repositories without all this verbose output on a "no change" scenario?

Andy

Andrew Schein <andrew@andrewschein.com> writes:

This is merely a historical accident that you cannot change. "git status" was originally a dual-purpose implementation detail of "git commit" command to:

 (1) report if there is anything you can commit with its exit code; and

 (2) if there is something to be committed, show what will and what won't
     be committed.

For the purpose of (1), the calling command "commit" was told by the user to make a commit, so the situation where there is nothing to commit was signalled as an error. The code conceptually looked like:

	(git-commit.sh)

	do the preparation
	if git-status
        then
        	# message "there is nothing to commit" already given
		revert the preparation
                exit 1
	fi
        really make a commit

These days, "git commit" has been rewritten and the logic to see if there is anything to commit is built-in, but people still use "git status" primarily for purpose (2), the listing of what's committed and what's left.

Oh, and your example is wrong. If you are going to commit with -a, and if you want to see if such a "git commit" has actually something to commit, then you should be giving -a to status as well, like:

	git status -a && git commit -a

Also, unless you ask explicitly, "git commit" will not make an empty commit, so you can always omit the "git status -a &&" part and have "git commit" itself do the checking.

Hi Junio -

Thanks for your reply giving historical context. The command line examples I gave were intended to give examples of the output rather than my own usage pattern.

Here is my actual usage pattern... from a file called "sync.sh" (would love feedback on whether this is the gitly? way to solve my use case):

# environment set up occurs before loop, I pull before "pushing" in an
attempt to prevent
# conflicts from being left on the shared repository.
for dir in ./* ; do
    if [ ! -d $dir ]      ; then continue ; fi #not a directory
    if [ ! -e $dir/.git ] ; then continue ; fi #not a git repo
    dir=`basename $dir`
    echo "syncing: $dir"
    set +e # commit returns an error if there is nothing to commit.
    (cd ./$dir ; git commit -a)
    set -e
    (cd ./$dir ; git pull $UP "$REPO/$dir" master)        #pull
    ssh $HOST "mkdir -p $DEST_CACHE/$LOC/$dir"  # these three lines
handle "push"
    rsync -rl --delete ./$dir/.git -e ssh "$DEST:$DEST_CACHE/$LOC/$dir/.git"
    ssh $DEST "(cd $LOC/$dir ; /tools/bin/git pull
$DEST_CACHE/$LOC/$dir/.git master)"

done

On Tue, May 12, 2009 at 12:24:56PM -0400, Andrew Schein wrote:

One trick to avoid playing with "set -e" is just:

  thing_whose_error_you_want_ignore || true

But that aside, it might be nice _not_ to ignore the result from "commit -a", since it could also be warning you that it failed to correctly commit. You probably want to do:

  git diff --quiet HEAD || git commit -a

That is "either there is nothing to commit, or we succeed in committing everything". That of course has a race condition between the two commands, though (i.e., if you modify files in between). If you want something truly atomic, I think we would need "git commit" to distinguish error codes between "I had nothing to commit" and "an error occurred" (returning "1" right now could mean either).

Pushing directly should be a lot more efficient in the case of repacking (git will realize that the remote already has certain objects, but if the filenames change, rsync will not).

-Peff

On Tue, 12 May 2009, Andrew Schein wrote:

What *is* your use case? What you're doing seems nuts to me (like, you're going to send out files with this script that someone is in the middle of editting), but I don't know what you're trying to do.

What are you trying to avoid by not using "git push"? Why are you committing whatever changes happen to be in working directories? Are you intending to be able to handle non-trivial merges?

	-Daniel
*This .sig left intentionally blank*

I am new to git... so my first instinct is to try to reproduce a work flow that I know works with mercurial setup. It is possible that the concepts don't translate correctly. Here goes...

I have a bunch of separate project-related repositories. There are very few users of the system. Most of the time I am the only user. I want a system for syncing my local repositories to a single shared repository. For example some days I work on my laptop, and some days from my desktop. A third "shared/public" repository "on campus" serves as an always available repository that anyone I collaborate with can pull from. Also it is backed up, and for this reason I designate it the "shared" version. So the purpose of the sync.sh script is to synchronize the personal laptop/desktop repository to the on-campus version.

Something I have learned from using mercurial in industry is that when somebody messes up a "public repo" with conflicts they frequently don't clean up the mess. This can be a sign that they have not learned the lessons of cleanliness rather than ill intent. Otherwise (and similarly) this messiness can be caused from not noticing that they have left a mess.

The motivation of having a sync script that is run on each user's local repository is to decrease the likelihood of a mess. This is achieved by first pulling from the common repository and resolving conflicts _before_ "pushing" (note quotations) their changes to the common repository. There is a possibility of a race condition that leaves a conflict on the shared repository, however the risk is diminished.

Finally, I use "push" in quotes because actually the script uses only uses the pull command. This prevents proliferation of branches on the shared repository.

Is there a better way to achieve this in git than the sync.sh script I sent around?

Thanks,

Andy

On Tue, 12 May 2009, Andrew Schein wrote:

In general, you should probably not have the script run "commit", but run it by hand whenever you've finished making changes (and probably done at least a quick syntax test, if possible). Otherwise, you'll often get the situation where you get distracted halfway through modifying a line, save something by habit, and then find this state getting sent to the system that other people pull from, and they'll find that they can't build your latest code.

In the normal process of pushing changes to a shared repository, there is no possibility of leaving conflicts, because there's no way to get git to attempt a merge on push; the push simply fails, and you then pull, resolve any conflicts, commit the merge, and try again at pushing.

It's better to configure the push refspec to do what you want, which seems to be "refs/heads/master:refs/heads/master", and use the fact that git push actually does what you're trying to do in a race-free and efficient way.

Also, the "refs/heads/*:refs/heads/*" refspec is only different in the case where you use branches other than "master", and seems like it would be useful for your use case; if you actually want to send everything you do to the shared location, you want to have all of the branches you've made. (That is, you get one branch on the shared repository per distinct name of a branch that you use on a local repository, not one "master" for each local repository's "master")

	-Daniel
*This .sig left intentionally blank*

On Tue, May 12, 2009 at 11:18:25AM -0400, Andrew Schein wrote:

Yes, I have such a script. I check:

  git ls-files -m -o -d --exclude-standard --directory --no-empty-directory

If it produces any output, then there is something to commit (either a change in a tracked file, or an untracked file that might need to be added).

I also do a fetch and check to see if we have any commits that need to be merged:

  git rev-list master..origin

or any commits that we need to push:

  git rev-list origin..master

(actually, it is a bit more complicated, since "master" and "origin" are just convention; I actually parse the config to find the branch pairs).

-Peff