Re: Managing sub-projects

On Sat, Jun 18, 2016 at 4:20 PM, Michael Eager <eager@eagerm.com> wrote:

I think forking is the way to go here, as you want to have new code and maintain that.

I assume you want to keep the history for each project separate, so I would recommend against using subtrees.

git slave looks interesting for your use case. (I looked at it once before, but have no deep knowledge about it as it is not part of core git)

IIUC the repo tool is tailored to be a multi-repo manager optimised for usage with Gerrit, not just plain Git. The repo tool tracks branches in the subproject instead of versions (as submodules do), so consistency is hard, specially when looking back in history. (Not sure if you care about that, but if you want to use e.g. git bisect, having an easy reproducable history is a must)

Personally I would try out submodules.

Beware that there are 2 areas you need to look at. First the submodule repo needs to have a remote that points away from the projects origin (to your private fork).

Then you have to look at the superproject that
1) records the sha1 for the submodules internally
2) all other information except the tracking sha1s must be user provided,
    where the .gitmodules file contains recommendations (i.e. the url where to
    obtain the submodule from, whether to clone it shallowly,
    if we have a specific branch in mind). The contents of that file
    are not binding, e.g. if the url provided in the .gitmodules file becomes
    outdated later, it is still possible to setup the
submodule/superproject correctly.

However for your business purpose, you would put the url of the private forks in the recommended URL of the submodules.

As the superproject only tracks the sha1, and has this recommended pointer where to get the submodule repository from, you need to take special care in a rebase workflow, because the old rebased commits fall out of the reachability of the graph of objects, e.g.:

Say you have a version `abc` in a submodule that is one commit on top of canonical projects history, and `abc` is recorded as the sha1 in the superproject.

Then you rebase the commit in the submodule to a newer version of the upstream, which then becomes a new commit `def` and `abc` is not referenced any more, so it can be garbage collected.

This is bad for the history of the superproject as it then points to an unreachable commit in its history.

To preserve the historic non-rebased `abc` commit, you could have a set of branches (or tags) that maintain all the old non rebased versions.

This problem comes up with submodules with any workflow that requires non fast forward changes (forced pushes), I think.

So maybe you need to have an alias in the submodule for rebasing, that is roughly:

rebase:
    if rebased history is published
        create a tag, e.g.: "$(date -I)-${sha1}"
        (and push that tag here or later?)
    rebase as normal
    carry on with life

To get back to your complaint:

I would imagine this to be

     (cd submodule && git remote set-url origin <your fork> && git push origin)

for plain pushing in the submodule and then

    $EDIT .gitmodules
    # edit submodule.<name>.url to point at <your fork>

to get the superproject correct.

Thanks, Stefan