Requirements for integrating a new git subcommand

I have completed work on git-weave (the tool I had called 'gitpacker' in some previous postings). I want to submit a patch that integrates it into git; in hopes of smoothing the process I have some technical and procedural questions.

First, however, let me present the git-weave documentation for design review:

---------------------------------------------------------------------- git-weave(1) ============

NAME ---- git-weave - Weave a sequence of trees and log data into a repository

SYNOPSIS -------- [verse] 'git-weave' [-v] [-m limit] [-q] indir outdir

DESCRIPTION ----------- git-weave accepts an input directory containing a sequence of subdirectories and a metadata file, and composes them into a git repository created under the specified output directory (which must not exist).

If the input directory is identifiably a git repository, the weave operation is reversed; tree states from each commit are unraveled into the output directory with a log holding commit metadata (committer/author/comment information and parent headers representing links of the repository DAG) and tags.

This tool is primarily intended for importing and working with project histories that have been preserved only as linear sequences of release snapshots. It may also be useful for surgery on linear repositories

While the weave operation can build a commit graph with any structure desired, an important restriction of the inverse (unraveling) operation is that it operates on *master branches only*. The unravel operation discards non-master-branch content, emitting a warning to standard error when it has to do so.

Commits from the repository's master branch are unraveled into directories named for integers from 1 increasing, but their order of composition when re-woven is actually set by the sequence of entries in the metadata file. File trees may be inserted or removed without hindering re-weaving provided the pointers in the log's parent fields are fixed up properly.

METADATA FILE FORMAT -------------------- The metadata file format will contain three kinds of stanzas: entries for commits, entries for lightweight tags, and entries for annotated tags.

A commit stanza has headers similar to those in a commit raw log: commit, committer, author, and optionally parent headers. The header contents are not hash IDs, but arbitrary text cookies either declared by a previous commit stanza or referencing one. The following example declares "8" to be a commit ID, and references a previous commit identified as '7'. Note that commit IDs are not required to be numeric strings, though the unravel operation generates them that way.

------------ commit 8 parent 7 author Eric S. Raymond <esr@thyrsus.com> 1325026869 +0000 committer Eric S. Raymond <esr@thyrsus.com> 1325026869 +0000

Initial revision . ------------

The text body of a commit comment or tag comment entry is delimited from the headers by an empty line; the text body must always end with "." on a line by itself; and text lines beginning with "." will have an additional "." prepended to them.

A commit stanza may also have a "directory" header. If present, this sets the name of the subdirectory in which git-weave expects to find the content tree for this commit. For example

------------ commit 24 directory intercal-0.17 parent 23 author Eric S. Raymond <esr@thyrsus.com> 1325026489 +0000 committer Eric S. Raymond <esr@thyrsus.com> 1325026489 +0000

The state of the INTERCAL project at release 0.17. . ------------

A label stanza declares a lightweight tag. This example declares a tag 'sample' pointing at the commit identified as 102.

------------ label sample refers-to 102 ------------

A tag stanza declares an annotated tag. This one declares a tag named 'annotated1' pointing at the commit declared as 99.

------------ tag annotated1 refers-to 99 tagger Eric S. Raymond <esr@thyrsus.com> Sat Nov 17 03:16:26 2012 -0500

This is an example annotated tag. . ------------

When you are composing commit and tag stanzas by hand, you can count on any of the date formats normally acceptable to git to be recognized.

If, when weaving, any committer or author or tagger line, the date is omitted, git-weave will supply as a default the latest modification time of any file in the corresponding tree.

If a committer or author or tagger line is omitted entirely, the user's name and email address as retrieved by ''git-config'' will be supplied as defaults, and the date will default as above.

Thus, the following variation on one of the previous examples is a valid stanza:

------------ commit 24 directory intercal-0.17 parent 23

The state of the INTERCAL project at release 0.17. . ------------

OPTIONS
-------
-q::
	Be quiet.  Suppress the normal spinning-baton progress meter
	with timing information.

-m::
	Limit the number of commits or trees processed to a specified
	integer maximum. '0' means process all of them.

-v::
	Be verbose, enabling progress and command-execution messages
	This option will probably be of interest only to developers;
	consult the source code for details.

EXAMPLES
--------
* Weave a sequence of trees in the directory 'unraveled' 
into a git repository in the directory 'repo'. 
+
------------
$ rm -fr repo; git-weave unraveled repo
------------
+
The metadata is expected to be in 'unraveled/log'.  This mode of
operation is triggered when there is no file 'unraveled/.git', 

* Unravel a repository in the directory 'repo' into a sequence
of file trees and a metadata log in the directory 'unraveled'.
+
------------
$ rm -fr unraveled; git-weave repo unraveled
------------
+
This mode of operation is triggered when there is a 'repo/.git' file.

SEE ALSO -------- linkgit:git-log[1] linkgit:git-checkout[1] linkgit:git-add[1] linkgit:git-mktree[1] linkgit:git-ls-tree[1] linkgit:git-update-references[1]

GIT --- Not yet part of the linkgit:git[1] suite ----------------------------------------------------------------------

Yes, there are scripts in contrib that do similar things. git-weave is an improvement is several ways: (a) it is documented, (b) I am shipping it with a functional test, (c) I am prepared to maintain it and am quite unlikely to drop out of sight :-), (d) it does both the import operation *and its inverse*, and (e) it is rather more powerful, including the ability to decorate the import with annotated tags.

Now *my* questions:

1. I have a round-trip test for the tool that I can very easily adapt
to speak TAP.  To function, the test will require a small linear
history to operate on in the form of an import-stream file (so the
result of round-tripping through a weave-unravel can be diffed against
the original).  Does the distribution include any test repos?  If
so, where can I find them?

2. I understand that a "git foo" command is typically implemented
as "git-foo" binary or script in /usr/lib/git-core.  What I don't
know is what the other interfacing requirements are.  Are they
documented anywhere?  In particular...

3. Is there any registration protocol other than simply installing 
the extension in the subcommand library?

4. How does "git help" work?  That is, how is a subcommand expected
to know when it is being called to export its help text?

5. I don't see any extensions written in Python.  Are there any special
requirements or exclusions for Python scripts?

On Wed, Nov 21, 2012 at 9:30 PM, Eric S. Raymond <esr@thyrsus.com> wrote:

...

No. We create the repositories from scratch using a series of commands. If you look at the test library the environment is set in a predictable way so the author, committer and timestamps are all set to a single well known value, allowing Git to create a commit that is reproducible on all platforms. A test_tick function is used in the scripts to move the clock, allowing different times to be used. For an example see t7502-commit.sh, or really any script in that directory.

Nope. "git foo" will invoke "git-foo" with GIT_DIR set in the environment, so you know what repository to act against, and so does any git command you recursively invoke. Other than that there really aren't any interface requirements. Your program is executed with whatever arguments the caller gave you. Its pretty simple UNIX stuff. :-)

Nope. Running "git whatever-this-is-i-have-no-idea" will try to execute "git-whatever-this-is-i-have-no-idea" via $PATH, after adding /usr/lib/git-core to the front of $PATH. Its pretty simple actually. If your standard C library can find the program in $PATH its run, if it can't find it, it dies.

IIRC "git help foo" runs "man git-foo".

Nope, it just has to be executable. We don't have any current Python code. IIRC the last Python code was the implementation of git-merge-recursive, which was ported to C many years ago. We avoid Python because it is not on every platform where Git is installed. Yes Python is very portable and can be installed in many places, but we prefer not to make it a requirement.

Shawn Pearce <spearce@spearce.org>:

OK, that makes sense.

I find that odd.  You avoid Python but use shellscripts?  *blink*

One would think shellscripts were a much more serious portability problem.

Eric S. Raymond venit, vidit, dixit 22.11.2012 23:11:

Different versions of python can be a mess to deal with, also, at least with respect to standard modules being "standard" or not for a specific version.

In any case, the point is that we try to avoid *additional* dependencies. Shell and perl are given with the status quo.

That being said, we also have remote helpers in python. The testsuite can run tests depending on the availability of python.

Regarding git-weave, I'm wondering (without having looked at the code) how this relates to git-archiv and git-fast-import/export, i.e. how much this leverages existing infrastructure rather than reinventing the wheel. Do your "trees" correspond to a "git tree"?

Again, without having looked at the code, it seems to me that exploding blob and tree objects might give you a structure not much unlike weave's, and your instruction sheet resembles that of fast-import quite a bit (plus date fill-in etc.).

One could even dream about implementing this as a remote helper instead...

Michael

Michael J Gruber <git@drmicha.warpmail.net>:

The unravel operation of git-weave is something like running git-archive on every revision and saving the results in sequentially-named directories, except that it also produces a metadata file that allows the operation to be inverted. So it is strictly more powerful.

The weave operation could be implemented using git fast-import, which I am quite intimately familiar with from having written reposurgeon. Functionally, the difference is that it would be a PITA to patch a fast-import stream to insert or modify or remove revisions in the middle, because the content of any given revision is in blobs that can stretch arbitrarily far back from its commit and are shared with other revisions. With git-weave tree-sequences these operations are easy and safe.

The weave log resembles an import stream, yes - that's because they
have to capture the same data ontology.  One major difference is that weave
logs are designed to be generated and edited by humans.
 

What is a "remote helper" in this context?

On Thu, Nov 22, 2012 at 11:11 PM, Eric S. Raymond <esr@thyrsus.com> wrote:

"git help" can also launch a browser to display the HTML version of the man page. It is looking for man pages and HTML docs in the paths given by "git --man-path" and "git --html-path".

Cheers, Christian.

Shawn Pearce <spearce@spearce.org>:

This turns out not to be quite true. The tree currently includes two Python scripts, a Perforce importer and a test helper.

I'm in he process of writing up a document on command integration based on your answers. I will submit it for incusion in the tree shortly.

Eric S. Raymond:

I was just about to say that the import direction of this seems to fill the same need as contrib/fast-import/import-directories.perl that I submitted a few years back.

Your version seems only to be able to import a linear history, however, my tool does support creating merge commits (basically, the history I had to import was very messy and contained a lot of snapshot directories having been worked on in parallel).

Hmm, indeed. I have been thinking of trying to wrap up the test suite I have locally into something that could work within the Git testing framework, but haven't found the time to or energy for so far.

Anyway, my sentiment is that if you can add support for merges in you weave tool, then I am very much for removing my old script from the repository.

Peter Krefting <peter@softwolves.pp.se>:

Yours was the closest in functionality, yes.

git-weave can *weave* (import) a sequence with merge commits, as your
tool can.  What it can't do is unravel a nonlinear repo into a tree
sequence that will round-trip through the weave operation.  (Though I
thought of a way I might be able to fix that last night.)
 

The support is there. I will take this as direction to (a) add a test load demonstrating this, and (b) include the removal of import-directories.perl in my integration patch.

Eric S. Raymond:

Yes, please.

"Eric S. Raymond" <esr@thyrsus.com> writes:

Imagine that I have a simple four-commit diamond:

    I---A
     \   \
      B---M

where Amy and Bob forked the initial commit made by Ian and created one commit each, and their branches were merged into one 'master' branch by a merge commit made by Mac. How many state snapshots will I see when I ask you to unravel this? Three, or four?

If you are going to give me all four states, then I do not understand why this needs to be limited to the master branch only. Even if you start from a single commit at the tip of 'master', once you hit a merge, you will need to follow all of two (or more) paths to dig down to the root(s), so supporting to start digging from more than one commit is not all that different.

If you are going to give me only three states, following the first parent ancestry chain, then the description needs to state it more clearly. I am not saying first-parent-only history is useless, but the user needs to know that merges are flattened in the unraveled result and the resulting history becomes linear, following the first parent ancestry chain of the original history (if that is what the tool does) before deciding if this tool matches what she needs.

As to the procedural stuff, I think others have sufficiently answered already. If I may add something, a new stuff typically start its life in contrib/ before it proves to be useful.

Junio C Hamano <gitster@pobox.com>:

You will see four tree states. I have managed to remove the master-branch-only restriction.

Thank you, I have submitted a documentation patch which folds in the on-list discussion.

As a separate point...are you requesting that I submit my integration patch to drop git-weave in contrib? If so, I will of course comply.

But I will point out that git-weave is not a half-thought out experiment; it is fully documented and has a functional test. The case for its usefulness is bolstered by one previous contrib script, which the author has agreed to retire in favor of git-weave.