Re: [RFC] Third round of support for cloning submodules

From: Junio C Hamano <junkio@cox.net>
Date: May 20, 2007, 19:10 UTC
Message-ID: <7vhcq7l3ar.fsf@assigned-by-dhcp.cox.net>
In-Reply-To: <11796842882917-git-send-email-skimo@liacs.nl>

skimo@liacs.nl writes:

> This patch series implements a mechanism for cloning submodules.

Let me start by asking a few stupid questions.

Show 6 quoted lines

> Each submodule is specified by a 'submodule.<submodule>.url'
> configuration option, e.g.,
>
> bash-3.00$ ./git-config --remote=http://www.liacs.nl/~sverdool/isa.git --get-regexp 'submodule\..*\.url' 
> submodule.cloog.url /home/sverdool/public_html/cloog.git
> submodule.cloog.url http://www.liacs.nl/~sverdool/cloog.git

You are priming the process by having these configuration variables in the superproject to be cloned (i.e. this is done by the owner of the superproject).

Show 18 quoted lines

> git-checkout will use the first url that works.
> E.g., a
>
> git clone --submodules ssh://liacs/~/public_html/isa.git
>
> followed by
>
> git checkout origin/submodule
>
> (which only works for me), will use the first url, while a
>
> git clone --submodules http://www.liacs.nl/~sverdool/isa.git
>
> followed by
>
> git checkout origin/submodule
>
> will use the second.

What's the design like to make clone and checkout work together? When you run the first clone with -n ("do not checkout"), what should happen (I am not asking what your code does, but what the desired behaviour should be)? My take on that question is "subproject cloning is done by checkout, not clone".

> The cloning of submodules is now handled inside git-checkout.

which I guess means your answer is the same as mine, which is fine.

I am very worried about this big red switch that says "all subprojects to be cloned and checked out, or nothing". I think this would not work well with projects that truly need superproject support (i.e. very large ones, where most people would not want to clone and check out every single subproject).

> I currently do not fetch after the initial clone, since
> I'm not sure what ref to use for the revision I need to
> fetch for the supermodule.

I think fetching inside the subproject can be safely done with the default (i.e. refs/heads/*:refs/remotes/origin/*) of 1.5.0 or later, as long as we tell the users of the feature that they should make sure that the commit referenced by superproject tree entries are available with such a fetch, which is a sane thing to require anyway.

The more important issue I think is at what point in the superproject operation does a recursive checkout in a subproject should happen, and how we should do the checkout. Issues I can think of offhand are (no way exhaustive):

 - Do we checkout a branch? if so which one?

 - Do we detach HEAD if the commit named by the superproject
   tree is not at the tip of the current branch of subproject?
   do we detach always even if the commit is at the tip?

 - What would we do when the subproject working tree is not
   clean?

 - How can a user decide which subproject to descend into and
   which subproject to ignore, and how does git remember the
   earlier decision made by the user without asking the same
   again, and how does a user express "now I want to also track
   that subproject I've ignored so far" and "now I am not
   interested in following that subproject anymore"?

So I tend to disagree with not having the indirection we discussed on the other thread about .gitmodules, but I consider it a minor detail of cloning, and it is not a major deal to me.

However, I agree with Alex that checkout semantics is a much bigger deal, and would expect people (brighter than myself, hopefully) to offer ideas.