My use case

My recent question about checking out an upstream commit (and ending up as a result on a detached head) has generated a lot of discussion and useful information. I thought I'd describe what I'm actually up to because someone might have some ideas on how to do it better.

I'm trying to integrate git into a Lisp IDE (Clozure Common Lisp). Because it's Lisp, code is often developed incrementally and experimentally. This has two important consequences. First, the files in your working directory can and often do, as part of the normal course of events, get into an incoherent state. Everything in your running Lisp image is working just fine, but if you try to recompile your entire system something will break, and you can go for a very long time without discovering that this has happened. Second, development can often lead to dead-ends where you want to throw everything you've just done away, go back to some earlier version of *some but not all* of your code, and start over.

In other words, it is not uncommon to want to roll back an individual file or set of files to an earlier version and leave the rest of the tree alone. git can do this, but it's not straightforward. Simply rolling back through the history of the current branch doesn't work because you might want to roll back file A, but the last dozen revisions or so have been changes to file B. You might also want to roll both A and B back to states which never co-existed in the original history.

One approach is to use git rev-list to find those commits where particular files changed, but this is sub-optimal for several reasons. First, a naive approach calls rev-list for every rollback, and rev-list has to traverse the entire history, so it's very inefficient. Second, if you roll back a single file, git doesn't keep track of that file's provenance, so you have to manually track which files and have been rolled back and which revisions they have been rolled back to. (There was a third problem but I can't think what it was right now.)

I have this intuition that git can be made to really do this right by keeping a separate history of every individual file in addition to a history of the entire source tree. Git can't do this directly as far as I know. I'd be writing additional code to generate extra tree and commit objects every time a file was saved from the IDE. But turning this intuition into reality is turning out to be quite challenging. So I'm going with the rev-list approach for the first version despite its shortcomings.

If anyone has ideas or suggestions, feedback would be much appreciated. But this is mostly just FYI.

rg

Excerpts from Ron Garret's message of Sat Jan 30 03:26:19 -0500 2010:

My approach, in this case, would be to use git log, possibly git log -p, in order to view the changes in each file you were interested in rolling back. If the rollback falls on a commit boundary, great; you can use `git checkout rev -- path`. If not, you can perform that, and then use git checkout -p to selectively revert hunks in the patch (fundamentally, you can't really get better than that).

While Git's ability to look at individual file's history is "slower", it's still quite excellent, and you are encouraged to use it as necessary.

Slightly relatedly, I'd recommend flushing to disk and committing more often. It's great that the LISP REPL allows for more daring changes, but having a history either way is very helpful!

Cheers, Edward

In article <1264840729-sup-5264@ezyang>,
 "Edward Z. Yang" <ezyang@MIT.EDU> wrote:

Don't forget, I'm integrating this *into* the IDE, not just using it *for* the IDE. So I want to just have a context menu on each code window with "SNAPSHOT" and "ROLLBACK" items that Just Work. The casual user won't even know that there's git behind the scenes.

Good to know.

Yep! Save early, save often. I'm actually considering an auto-save option where it takes a snapshot every time you evaluate a form after making a change.

rg

On Sat, Jan 30, 2010 at 12:54:13AM -0800, Ron Garret wrote:

This is a workflow question, I suppose, but I find things work much better if you can get the user to give you explicit commit boundaries so that (a) bisect works, and (b) they can describe what each commit does, and (c) so they can more easily move specific bug fixes or features between different release branches. The free-form hacking more may be nice, and very "LISP-like", but there are some real advantages to having explicitly describable and documented commits.

Best regards,

					- Ted

In article <20100130174844.GD788@thunk.org>, tytso@mit.edu wrote:

You are absolutely right. That is another reason why having the individual files tracked separately from the main project would be a good thing if I can get it to work. (It would be kind of like having a git-stash on a per-file basis.)

rg

Ron Garret <ron1@flownet.com> writes:

When you have more than one functions defined in a file, and the interactive Lisp development cycle works primarily on s-exp basis, not necessarily constrained by file boundaries, don't you want even finer grained control than "stash per-file"?

I don't think of a good solution myself, but I find your "finer than whole tree" an interesting topic.

In article <7vtyu3o1hq.fsf@alter.siamese.dyndns.org>,
 Junio C Hamano <gitster@pobox.com> wrote:

You're reading my mind. My long-term vision for this thing is indeed a revision control system based on S-expressions, not files. (I personally think the whole concept of files is fundamentally broken, but that's a whole 'nuther kettle of lobster.) Unfortunately, that opens up a Pandora's box of ancillary issues, which would be seriously off-topic here. If you're interested we can talk about it off-line (or we can do it here if you think it's not too much of a tangent).

rg