[RFC] New commit object headers: generation and note headers

As new major git release 1.6.0 is close (BTW. I wonder if git would ever reach/get 2.0.0 release...), I'd like to sum up here, adding my own thoughts and comments, ideas about extending commit object by adding new headers. I think it would be better to have such major feature introduced in major release, and not with only minor number changed. For some headers the faster it is introduced the better.

1. 'generation' header

In the "[BUG?] git log picks up bad commit" thread:
  http://permalink.gmane.org/gmane.comp.version-control.git/72274
later "[RFH] revision limiting sometimes ignored" there was resurrected 
idea of the 'generation' header. This header is meant to simplify 
removing uninteresting commits in the presence of clock skew, to 
replace various commit-time related heuristics.

The proposed solution (which was at least once discussed in the past on 
git mailing list) is to use for this "generation number":
 1. For parentless (root) commits it equals 1 (or 0)
 2. For each commit, it equals maximum of generation numbers of parents,
    plus 1.
Of course to not to have to recalculate it from beginning it must be 
saved somewhere. Best solution is to use 'generation' header for that.

Unfortunately there is complication that commits written before this header introduced doesn't have generation number handy. It was proposed then to use generation number if possible, and fallback to old date based heuristic if it does not exist, and do not (re)calculate it; the idea is to avoid such cost.

My comments: ============ The problem is twofold: when to calculate generation header, and what to do with commits that lacks it. We could require to calculate generation header when creating a commit (commit, amend, rebase, filter-branch), but this might mean that a few first commits after 'generation' header is introduced would be much slower.

As for older commits which lacks generation number header: perhaps some (pack)-index-like external storage/cache, where generation numbers will be saved as we generate them? And perhaps some command to generate generation numbers in advance, in a free time.

Note that keeping generation numbers externally to the object database is more error prone (cache sync), and would not propagate.

The question is if to take grafts and shallows when creating version numbers: if they are to be saved in object database, then no. If saving to external pack-index like storage, then perhaps.

2. 'note' header (no semantical meaning)

There was some time ago discussion about adding 'note' header, initially to save original sha-1 of a commit for cherry-picking and rebase; then for saving explicit rename or corrected rename info, for saving chosen merge strategy, and for saving original ID of SCM import.

My comments: ============

for a commit, for commits imported from other SCM. This way we do not have to parse commit message (fragile and ugly, and makes it harder for two-way exchange: no pristine commit message), or store them externally (not propagated, prone to be lost).

Another would be to save rename and copy info when importing from another SCM which tracks renames and not detects code movement. This would allow (at least theoretically) for lossless import. When detecting renames, in the process of finding common merge base(s), we could check and take into account such information. It would be purely advisory...

On Sat, 9 Feb 2008, Jakub Narebski wrote:

Surely, at least sombody has to do the slow commit that's the first one with a generation number. Maybe make it optional? If Linus calculates the generation number of some v2.6.x and only merges trees that have been rebased on it by people with sufficiently recent git, it should only take a lot of time once.

It's probably best to start by only including them if all parents have them, then having a cycle where figuring them out slowly is optional and defaults to off, and then make it default to on once projects are likely to have them in general, or at least would be likely to retain them once a few slow commits are done.

Probably want to have a prescribed syntax for specifying what note this is, so that different programs using notes don't confuse each other.

	-Daniel
*This .sig left intentionally blank*

On Sun, Feb 10, 2008 at 12:35 AM, Daniel Barkalow <barkalow@iabervon.org> wrote:

How about git ignoring all X- headers and let programs freely add them? For example, X-SVN may be used for git-svn.

"Nguyen Thai Ngoc Duy" <pclouds@gmail.com> writes:

Please don't.

When two people/programs create an otherwise identical (for the purpose of git) commits that have two different object names, there'd better be a very good reason other than "I felt like adding an extra header that I can use willy-nilly".

Please separate the 'note' part and the 'generation' part and make two separate discussion threads.

And kill 'note' part altogether, but that can be done in that thread ;-).

On Sat, 9 Feb 2008, Junio C Hamano wrote:

Ah, well... very good reservation. So here it goes generic 'note' (or 'X-*', similarly to non-standarized email headers) header...

Still I think it is would be nice to have original commit id in a header when importing from foreign SCM. First, it would not pollute commit message, which would be identical with the original commit message (which allows easy two-way interaction). Second, it is much easier and much less error prone to extract it by machine.

As to marking explicitely renames and copies...

Hi,

On Sun, 10 Feb 2008, Jakub Narebski wrote:

I cannot agree to either reason. It is _not_ a git specific header, so it does not belong in the commit header.

Also, I find it does not clutter the commit message _at all_, but adds information that the user might find useful.

Lastly, I cannot see _any_ reason why it should be _easier_ or _less error prone_ to put an "original commit id" into the commit header than into the commit body.

Ciao, Dscho

On Sunday, 10 February 2008, Johannes Schindelin wrote:

Well, that, and the fact that the same commit imported using two different tools, one using this header and one didn't would result in different commit object... although if they differ in adding original revision id to the commit message commit objects would differ too.

Revisions ids can be long, and together with prefix introducing original SCM revision identifier be longer than customary 80 characters.

Besides, "git cherry-pick" was changed to _not_ add information about original commit id by default. Shouldn't this apply also for import?

Well, if commit message talks about foreign commit IDs...