patch, 8 partsgit-repack --max-pack-size

The three most common ways of making large packfiles are git-fast-import, the first git-repack, or git-repack -a. The first already supports a "--max-pack-size=N" option, which limits the resulting packfiles to N megabytes. This patchset adds the same option, with the same behavior, to git-repack to handle the other two cases.

This iteration fixes miscellaneous issues discussed on the list and introduces no behavior not seen elsewhere in git. It is based on "next" in order to incorporate Nicolas Pitre's work.

Dana How <danahow@gmail.com> writes:

Heh, since last night, 'master' and 'next' points at exactly the same tree object, although their commit ancestry are vastly different for obvious reasons.

Which leaves 'master' right now at v1.5.2-rc1 while 'next' at v1.5.2-rc1-687-gcb3892c; we might want to do something about this apparent discrepancy.

Junio C Hamano <junkio@cox.net> wrote:

Are you considering rewinding next to master? I'm not sure how else we could cancel out all of those merges and reverts of bad ideas from next.

One option might be to teach git-describe to look for matching trees, and see if git-describe could match the tree that v1.5.2-rc1 points at with the tree that cb3892c points and, and just call the latter the former... But that sounds a little risky to me in some projects, I'm not convinced that is always the right way to describe a history.

On Tuesday 2007 May 01, Junio C Hamano wrote:

It's perfect - I'd say that it's exactly right.

git-describe is for making unique - human readable names for points in history, not for describing the tree. It makes no difference that A and B have the same tree, they are different points.

I've always thought of git-describe as being a way of mapping a commit hash to a nicer looking name. If A and B are different commits then they should have different names.

Andy

Andy Parkins <andyparkins@gmail.com> writes:

Although 'next' and 'master' have vastly different histories behind them, among 685 commits between master and next, only 27 are non-merges, among which 20 are "real" commits that tried to advance various topics that failed, and 7 are reverts against these 20 commits [*1*].

Even though hstory of failed attempts (why they initially seemed good ideas, how they tried to solve problems, and why they turned out not to be so good ideas in the end) are interesting, if we ignore the failed attempts, iow, if we view the history from the point of view of "surviving features", the development history of 'next' and 'master' are moral equivalents. Yes, most of them were merged way earlier in 'next' than 'master', many of them were merged in multiple steps of two and three to 'next' and then finally merged to 'master' with a single merge, but the changes did hit both 'master' and 'next' (obviously, that is why we ended up with the same trees).

So the equivalence of 'master' and 'next' tonight is not quite the same as equating two random commits that happen to have the same tree. v1.5.2-rc1 and v1.5.2-rc1-687-gcb3892c should naturally have the same tree because they share conceptually the same history.

But I was not talking about changing describe output because of the above argument. What I was wondering was that it might be a good idea to loosen the promise of never rewinding 'next'. It might be easier to view the history of 'next' during development for each cycle, if it started afresh after a feature release.

Since now we are in a stabilization freeze, I expect that 'master' and 'next' will always have identical trees until v1.5.2 final. We _could_ declare now that 'next' will be reset to 'master' when v1.5.2 happens, and people who forked from 'next' to do their own customization can rebase any time that is convenient for them between tonight and v1.5.2 final.

I was not sure if that is even a good idea, and I am now inclined to think that keeping the failed attempt history is probably better than potentially causing confusion to people who follow 'next'. But it _is_ a possibility to reset 'next' to 'master'.

[Footnote]

*1* Some topics I did not use "git revert" to revert them one-by-one; instead, I reverted a whole topic with one commit.

On Tuesday 2007 May 01, Junio C Hamano wrote:

This is an interesting philosophy-of-version-control question. If two identical trees fall in the forest and there is no one there to diff them, was a release made? :-)

It's been my experience that failed attempts and dead-end branches are often of equal value to the successful branches. It's very handy when someone asks "why can't we do it like this", to be able to answer "look at revision xyz onwards". Even just for your own reference, I've often looked back on abandoned paths and thought "that wasn't as bad as I thought, I just need to fix it here and here" - if I'd discarded that work it would be gone forever.

It's certainly true in academia, a large part of my doctoral thesis was about "things that don't work" :-) Documenting failure is as important as documenting success.

Andy

On Tue, 1 May 2007, Junio C Hamano wrote:

And what is the advantage of doing that, exactly?

Nicolas

Nicolas Pitre <nico@cam.org> writes:

Not much and that is why I am not sure if that is a good idea, but one thing is:

	$ git log master..next
	$ git log --no-merges master..next

would start showing what are still pending in next without noise.

With the current 'next', the output of the former is almost useless, and the latter is useful only until 'next' starts diverging again from the master and if you are interested in what will never merged to 'master'. Once 'next' starts growing again, you will get mixture of what's cooking and what's killed; the output from the latter would become useless again.

On Tue, 1 May 2007, Junio C Hamano wrote:

If you do so, then please tag the branch before resetting it. I think it contains valuable data that would get lost otherwise.

Nicolas

Hi,

On Tue, 1 May 2007, Junio C Hamano wrote:

Of course, you could graft a fake relationship there. Something like

echo $(git rev-parse master master^ next) >> .git/info/grafts

should do. And you can do it _locally_, since all you care about is the higher-level relationship that _you_ know about. And this is the way to teach Git about it.

Ciao, Dscho

Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:

In case you did not get it, making the above usable is not about _me_.