Handling renames.

I've been looking at tracking file revisions. One proposed solution was to have a separate revision history for individual files, with a new kind of 'filecommit' object which parallels the existing 'commit', referencing a blob instead of a tree. Then trees would reference such objects instead of referencing blobs directly.

I think that introduces a lot of redundancy though, because 99% of the time, the revision history of the individual file is entirely reproducible from the revision history of the tree. It's only when files are renamed that we fall over -- and I think we can handle renames fairly well if we just log them in the commit object.

My 'gitfilelog.sh' script is already capable of tracking a given file back through multiple tree commits, listing those commits where the file in question was actually changed. It uses my patched version of diff- tree which supports 'diff-tree <TREE_A> <TREE_B> <filename>' in order to do this.

By storing rename information in the commit object, the script (or a reimplementation of a similar algorithm) could know when to change the filename it's looking for, as it goes back through the tree. That ought to be perfectly sufficient.

So a commit involving a rename would look something like this...

	tree 82ba574c85e9a2e4652419c88244e9dd1bfa8baa
	parent bb95843a5a0f397270819462812735ee29796fb4
	rename foo.c bar.c
	author David Woodhouse <dwmw2@hades.cambridge.redhat.com> 1113499881 +0100
	committer David Woodhouse <dwmw2@hades.cambridge.redhat.com> 1113499881 +0100
	Rename foo.c to bar.c and s/foo_/bar_/g

Opinions? Dissent? We'd probably need to escape the filenames in some way -- handwave over that for now.

On Thu, 14 Apr 2005, David Woodhouse wrote:

Please don't. It's fundamentally the git notion of "content determines objects".

It also has no relevance. A "rename" really doesn't exist in the git model. The git model really is about tracking data, not about tracking what happened to _create_ that data.

The one exception is the commit log. That's where you put the explanations of _why_ the data changed. And git itself doesn't care what the format is, apart from the git header.

So, you really need to think of git as a filesystem. You can then implement an SCM _on_top_of_it_, which means that your second suggestion is not only acceptable, it really is the _only_ way to handle this in git:

Except I want that empty line in there, and I want it in the "free-form" section. The "rename" part really isn't part of the git header. It's not what git tracks, it was tracked by an SCM system on top of git.

So the git header is an "inode" in the git filesystem, and like an inode it has a ctime and an mtime, and pointers to the data. So as far as git is concerned, this part:

	tree 82ba574c85e9a2e4652419c88244e9dd1bfa8baa
	parent bb95843a5a0f397270819462812735ee29796fb4
	author David Woodhouse <dwmw2@hades.cambridge.redhat.com> 1113499881 +0100
	committer David Woodhouse <dwmw2@hades.cambridge.redhat.com> 1113499881 +0100

really is the filesystem "inode". The rest is whatever the filesystem user puts into it, and git won't care.

The fact that git handles arbitrary filenames (stuff starting with "." excepted) doesn't mean that the SCM above it needs to. Quite frankly, I think an SCM that handles newlines in filenames is being silly. But a _filesystem_ needs to not care.

There are too many messy SCM's out there that do not hav ea "philosophy". Dammit, I'm not interested in creating another one. This thing has a mental model, and we keep to that model.

The reason UNIX is beautiful is that it has a mental model of processes and files. Git has a mental model of objects and certain very very limited relationships. The relationships git cares about are encoded in the C files, the "extra crap" (like rename info) is just that - stuff that random scripts wrote, and that is just informational and not central to the model.

		Linus

On Thu, 2005-04-14 at 11:11 -0700, Linus Torvalds wrote:

Note that not only may you have a _set_ of renames, but you'll also have a _different_ set of renames for each parent. Consider the representation of a merge where a file was called 'foo' in one parent, 'bar' in the other, and we called it 'foobar' in the resulting tree.

That's the main reason I wanted the renames in with the parent information -- so it's <parent><rename><rename><...><parent><rename>...

I see your point though and I can't be bothered to argue for the sake of the slight efficiency benefit we might gain from doing it that way. The implementation details really aren't that interesting right now.

Let us assume, however, that we have this information somehow stored in each commit object. It's perfectly sufficient from the POV of the 'git revtool' which I've been poking at; is it good enough for merges?

Consider a simple case: A branch is taken, file foo.c is renamed to bar.c, and now we're trying to merge that branch back into the head, which has moved on.

We can't just take 'bar.c' as a new file -- we have to track it all the way back to its inception, and notice that it actually shares a common ancestor with 'foo.c' in the other parent of the merge.

How feasible, and how computationally expensive, is that task going to be? Especially given that there may be _many_ new files that we need to attempt to tie up with their partners, across many potential renames.

One option for optimising this, if we really need to, might be to track the file back to its _first_ ancestor and use that as an identification. The SCM could store that identifier in the blob itself, or we could consider it an 'inode number' and store it in git's tree objects.

If we can avoid that, however, it would be nice. How feasible is the merge going to be without it?

* David Woodhouse <dwmw2@infradead.org> wrote:

how about the following structure:

    - tree_new --->
    - tree_old ---> rename_commit -> blob

the rename_commit object just contains a pointer to the file content blob. If a rename happens then the old tree references the rename_commit object (instead of the blob), and the new tree references it too. This way there's no need to list the rename via namespace means: if a tree entry points to a rename_commit object then a rename happened and the rename_commit object is looked up in the old tree to get the old name.

there's no redundancy caused by this method: only renames (which are rare) go through the rename_commit redirection. (to speed up the lookup the rename_commit object could cache the offset of the two names within their tree objects.)

	Ingo

On Thu, 14 Apr 2005, Ingo Molnar wrote:

Bzzt. Wrong.

ANYTHING you do with games like this will cause the "same directory creates different object" crap.

Git doesn't do that. There fundamentally is no history in objects, _except_ for the commit object. Two objects with the same name are identical, and that means that they are easy to share.

Any time you break that model, you break the whole point of git. Don't do it. You'll be very very sorry if you ever do, because it breaks the clean separation of "time" and "space". I guarantee you that your merges will become _harder_ rather than easier.

What you can do at an SCM level, is that if you want to track renames, you track them as a separate commit altogether. Ie if you notice a rename, you first commit the rename (and you can _see_ it's a rename, since the object didn't change, and the sha1 stayed the same, which in git-speak means that it is the same object, ie that _is_ a rename as far as git is concerned), and then you create the "this is the data that changed" as a _second_ commit.

But don't make it a new kind of commit. It's just a regular commit, dammit. No new abstractions.

Trust me, it's worth it to follow the rules. You don't start making up new concepts for every new thing you track. Next you'll want "tag objects". That's a totally idiotic idea. What you do is you tag things at a higher level than git ever is, and git will _never_ have to know about tag objects.

Some "higher level" thing can add its own rules _on_top_ of git rules. The same way we have normal applications having their _own_ rules on top of the kernel. You do abstraction in layers, but for this to work, the base you build on top of had better be damn solid, and not have any ugly special cases.

		Linus

* Linus Torvalds <torvalds@osdl.org> wrote:

ok, i accept your point of not putting this into such a low level as the object abstraction. Was a bad idea.

but i dont think the above would be enough: there can be renames of objects that have the same sha1 hash as other objects in the same tree, and developers want to track individual objects, regardless of whether other files share the same content. So some formal operation would be needed to signal renames - e.g. to embedd it in the commit object, per David's suggestion.

The thing i tried to avoid was to list long filenames in the commit (because of the tree hierarchy we'd need to do tree-absolute pathnames or something like that, and escape things, and do lookups - duplicating a VFS which is quite bad) - it would be better to identify the rename source and target via its tree object hash and its offset within that tree. Such information could be embedded in the commit object just fine. Something like:

me bb95843a5a0f397270819462812735ee29796fb4 tree 1756b578489f93999ded68ae347bef7d6063101c parent 9f02d4d233223462d3f6217b5837b786e6286ba4 author committer rename 39021759c903a943a33a28cfbd5070d36d851581 15234 9f02d4d233223462d3f6217b5837b786e6286ba4 16163

?

	Ingo

On Thu, 2005-04-14 at 20:58 +0200, Ingo Molnar wrote:

Actually I'm not sure that's true. Let's consider the two main users of this information.

Firstly, because it's what I've been playing with: to list a given file's revision history, I currently work with its filename -- walk the commit objects, inspecting the tree and selecting those commits where the file has changed. If my filename is 'fs/jffs2/inode.c' then I can immediately skip over a commit where the 'fs' entry in the top-level tree is identical to that in the parent, or I can skip a commit where the 'jffs2' entry in the 'fs' subtree is identical to the parent... it's all done on filename, and the {parent, entry} tuple wouldn't help much here; I'd probably have to convert back to a filename anyway.

Secondly, there's merges. I've paid less attention to these (see mail 5 minutes ago) but I think they'd end up operating on the rename information in a very similar way. To find a common ancestor for a given file,, we want to track its name as it changed during history; at that point it's all string compares.

Linus Torvalds wrote:

Maybe you (or the group) should standardize on a way to 'extend' the commit 'object' in terms of:

the layer1 (git) header for commit object is defined as such-and-such the layer2 (scm or other) header for commit object is defined as such-and-such

Much the way network protocols stack on top of each other. If a standard way of stacking is defined, then it could be much cleaner for future implementors to understand a 'new' stacking protocol, and it will make the scm-level extensions easier to discuss it terms of their own 'layer'.

David

David Woodhouse wrote:

For readability and simplicity I'd suggest using either URL-style %XX escapes or octal \xxx escapes for anything bytes < 33, minus the escape character.

Although Linus is correct in that an SCM doesn't *have* to handle this, it really feels like shooting for mediocracy to me. We might as well design it right from the beginning.

	-hpa

On Thu, 14 Apr 2005, H. Peter Anvin wrote:

No. git is not an SCM. it's a filesystem designed to _host_ an SCM, and that _is_ doing it right from the beginning.

Keep the abstractions clean. Do _not_ get confused into thinking that git is an SCM. If you think of it that way, you'll end up with crap you can't think about.

And at a filesystem layer, "rename" already exists. It's moving an object to a new name in a tree. git already does that very well, thank you very much.

But a filesystem rename is _not_ the same thing as an SCM rename. An SCM rename is built on top of a filesystem rename, but it has its own issues that may or may not make sense for the filesystem.

		Linus

Linus Torvalds wrote:

I wasn't referring to git per se, I was referring to the hosted SCM.

	-hpa

Linus Torvalds wrote:

I imagine quite a few folks expect something not entirely unlike an SCM to emerge from these current efforts. Moreover, Petr's 'git' scripts wrap your "filesystem" plumbing to that very end.

To avoid confusion, I think it would be better to distinguish the two layers, perhaps by calling the low-level plumbing... 'gitfs', of course.

Cheers,

Zach Welch Superlucidity Services

Zach Welch pontificated:

Or perhaps to come up with a name (or at least nickname) for the SCM.

GitMaster?

On Thursday 14 April 2005 01:40 pm, Andrew Timberlake-Newell wrote:

Cogito.  "Git inside" can be the first slogan.

Differentiating the SCM built on top of git from git itself is probably worthwhile to avoid confusion. Other SCMs may be developed later, built on git, and these can come up with their own clever names.

Steven

Dear diary, on Thu, Apr 14, 2005 at 10:42:16PM CEST, I got a letter where Steven Cole <elenstev@mesatop.com> told me that...

What about tig?

Petr Baudis wrote:

I like "Cogito"; it's a real name, plus it'd be a good use for the otherwise-pretty-useless two-letter combination "cg".

	-hpa

Dear diary, on Thu, Apr 14, 2005 at 10:58:52PM CEST, I got a letter where "H. Peter Anvin" <hpa@zytor.com> told me that...

Duh, believe me or not but I completely missed the "Cogito" part of Steven's mail. Of course, I like it too.

I'll commit my poor man's git-merge-in-separate-tree and finally get some sleep. I promise.

Steven Cole wrote:

And the logo could be a dove which, as everybody knows, coos.

Peter

On Thu, 14 Apr 2005, David Woodhouse wrote:

I personally think renames are a minor thing that doesn't happen much. What actually happens, in my opinion, is that some chunk of a file is moved to a different, possibly new, file. If this is supported (as something that the SCM notices), then a rename is just a special case where the moved chunk is a whole file.

I think that it should be possible to identify and tag "big enough" deletions and insertions, and compare them to find moves, where a further change may be applied in the middle if two chunks are "very similar" but not the same.

On the other hand, I think that the SCM will need to cache its understanding of what a commit did in order to give reasonable performance for operations like "annotate", and it may be advantegous to distribute things from this cache, since the committer might want to tell the system something that it didn't guess.

At some point, I'm going to argue for core support for "back pointers", where a file can be created which is "about" some other file(s), and someone looking for files "about" a particular file can find them without searching the entire database. I think this will turn out to be important for a variety of cases where some later participant wants to say something about an existing file without changing the content of the file.

	-Daniel
*This .sig left intentionally blank*

On Thu, 2005-04-14 at 18:23 -0400, Daniel Barkalow wrote:

Certainly we'd discussed the possibility that the 'rename' field may contain more than one destination, or more than one source filename. This could happen when a file is split into two, or when two files are merged into one, for example.