Consider two use cases:

  1. Using a v6 repo with locked files on a crippled filesystem not supporting symlinks. For the files to be usable, they need to be unlocked. But, the user may not want to unlock the files everywhere, just on this one crippled system.
  2. ?hide missing files

Both of these could be met by making git-annex sync maintain an adjusted version of the original branch, eg adjusted/master.

There would be a adjustment function. For #1 above it would simply convert all annex symlinks to annex file pointers. For #2 above it would omit files whose content is not currently in the annex. Sometimes, both #1 and #2 would be wanted. The function is currently implemented as the git-annex-adjust command.

[Alternatively, it could stay on the master branch, and only adjust the work tree and index. See WORKTREE notes below for how this choice would play out.]

adjusting

master           adjusted/master
A
|--------------->A'
|                |

When generating commit A', reuse the date of A and use a standard author, committer, and message. This means that two users with the adjusted branch checked out and using the same adjustments will get identical shas for A', and so can collaborate on them.

commit

When committing changes, a commit is made as usual to the adjusted branch. So, the user can git commit as usual. This does not touch the original branch yet.

Then we need to get from that commit to one with the adjustments reversed, which should be the same as if the adjusted branch had not been used. This commit gets added onto the original branch.

So, the branches would look like this:

master           adjusted/master
A
|--------------->A'
|                |
|                C (new commit)
B < - - - - - - -
|                
|--------------->B'
|                |

Note particularly that B does not have A' or C in its history; the adjusted branch is not evident from outside.

Also note that B gets adjusted and the adjusted branch is rebased on top of it, so C does not remain in the adjusted branch history either. This will make other checkouts that are in the same adjusted branch end up with the same B' commit when they pull B.

There may be multiple commits made to the adjusted branch before any get applied back to the original branch. This is handled by reverse adjusting commits one at a time and rebasing the others on top.

master           adjusted/master
A
|--------------->A'
|                |
|                C1
|                |
|                C2


master           adjusted/master
A
|--------------->A'
|                |
|                C1
B1< - - - - - - -
|
|--------------->B1'
|                |
|                C2'
B2< - - - - - - -
|
|--------------->B2'

[WORKTREE: A pre-commit hook would be needed to update the staged changes, reversing the adjustment before the commit is made. All the other complications above are avoided.]

merge

This would be done by git annex merge and git annex sync, with the goal of merging origin/master into master, and updating adjusted/master.

Note that the adjusted files db needs to be updated to reflect the changes that are merged in, for object add/remove to work as described below.

When merging, there should never be any commits present on the adjusted/master branch that have not yet been propigated back to the master branch. If there are any such commits, just propigate them into master before beginning the merge. There may be staged changes, or changes in the work tree.

First, merge origin/master into master. This is done in a temp work tree and with a temp index, so does not affect the checked out adjusted branch.

(Note that the reason this is done, rather than adjusting origin/master and merging it into the work tree, is that merge conflicts would be very common with the naive approach, because the adjusted branch often changes files, and origin/master may change the same files.)

origin/master  master   adjusted/master
A------------->A- - - ->A'
|              |
B------------->C

While a fast-forward merge is shown here, other merges work the same way. There may be merge conflicts; if so they're auto-resolved.

Then, adjust merge commit C, and merge that into adjusted/master.

origin/master  master   adjusted/master
A------------->A- - - ->A'
|              |        |
B------------->C- - C'->D'

This merge is done in-worktree, so the work tree gets updated. There may be more merge conflicts here; they're also auto-resolved.

Now, D' is a merge commit, between A' and C'. To finish, change that commit so it does not have A' as its parent.

This can be accomplished by propigating the reverse-adjusted D' back to master, and then adjusting master to yield the final adjusted/master.

origin/master  master   adjusted/master
A------------->A
|              |         
B------------->C         
               |         
           D - - -> D'

Notice how similar this is to the commit graph. Indeed, "fast-forward" merging the same B commit from origin/master will lead to an identical sha for B' as the original committer got!

Since the adjusted/master branch is not present on the remote, if the user does a git pull, it won't merge in changes from origin/master. Which is good because the adjustment needs to be applied first.

However, if the user does git merge origin/master, they'll get into a state where the adjustment has not been applied. The post-merge hook could be used to clean up after that. Or, let the user foot-shoot this way; they can always reset back once they notice the mistake.

[WORKTREE: git pull would update the work tree, and may lead to conflicts between the adjusted work tree and pulled changes. A post-merge hook would be needed to re-adjust the work tree, and there would be a window where eg, not present files would appear in the work tree.]

annex object add/remove

When objects are added/removed from the annex, the associated file has to be looked up, and the adjustment applied to it. So, dropping a file with the missing file adjustment would cause it to be removed from the adjusted branch, and receiving a file's content would cause it to appear in the adjusted branch. TODO

These changes would need to be committed to the adjusted branch, otherwise git diff would show them.

How to avoid making a new commit each time a single object is added/removed? That seems too expensive in both CPU and dangling git objects for old versions of the adjusted branch. It would be fine if git annex get and git annex drop only re-adjusted the branch one time at the end. OTOH, when should the assistant re-adjust the branch?

Maybe instead of re-adjusting the branch after each file, stage the worktree change, and hold off on committing. Then when a commit is eventually made, the reverse adjusting to propigate it to master would need to make sure to not remove files that were deleted as part of the commit, if their content is not present.

[WORKTREE: Simply adjust the work tree (and index) per the adjustment.]

reverse adjusting commits

A user's commits on the adjusted branch have to be reverse adjusted to get changes to apply to the master branch.

This reversal of one adjustment can be done as just another adjustment. Since only files touched by the commit will be reverse adjusted, it doesn't need to reverse all changes made by the original adjustment.

For example, reversing the unlock adjustment might lock the file. Or, it might do nothing, which would make all committed files remain unlocked.

push

The new master branch can then be pushed out to remotes. The adjusted/master branch is not pushed to remotes. git-annex sync should automatically push master when adjusted/master is checked out.

When push.default is "simple" (the new default), running git push when in adjusted/master won't push anything. It would with "matching". Pity. (I continue to feel git picked the wrong default here.) Users may find that surprising. Users of git-annex sync won't need to worry about it though.

[WORKTREE: push works as usual]

acting on filtered-out files

If a file is filtered out due to not existing, there should be a way for git annex get to get it. Since the filtered out file is not in the index, that would not normally work. What to do?

Maybe instead of making a branch where the file is deleted, it would be better to delete it from the work tree, but keep the branch as-is. Then git annex get would see the file, as it's in the index.

But, not maintaining an adjusted branch complicates other things. See WORKTREE notes throughout this page. Overall, the WORKTREE approach seems too problematic.

Ah, but we know that when adjustment #2 is in place, any file that git annex get could act on is not in the index. So, it could look at the master branch instead. (Same for git annex move --from and git annex copy --from and the assistant.)

OTOH, if adjustment #1 is in place and not #2, a file might be renamed in the index, and git annex get $newname should work. So, it should look at the index in that case.

problems

Using git checkout when in an adjusted branch is problematic, because a non-adjusted branch would then be checked out. But, we can just say, if you want to get into an adjusted branch, you have to run git annex adjust Or, could make a post-checkout hook. This is would mostly be confusing when git-annex init switched into the adjusted branch due to lack of symlink support.

After a commit to an adjusted branch, git push won't do anything. The user has to know to git-annex sync. (Even if a pre-commit hook propigated the commit back to the master branch, git push wouldn't push it with the default "matching" push strategy.)

Tags are bit of a problem. If the user tags an ajusted branch, the tag includes the local adjustments.
[WORKTREE: not a problem]

If the user refers to commit shas (in, eg commit messages), those won't be visible to anyone else.
[WORKTREE: not a problem]

When a pull modifies a file, its content won't be available, and so it would be hidden temporarily by adjustment #2. So the file would seem to vanish, and come back later, which could be confusing. Could be fixed as discussed in deferred update mode. Arguably, it's just as confusing for the file to remain visible but have its content temporarily replaced with a annex pointer.

master push overwrite race (fixed)

There are potentially races in code that assumes a branch like master is not being changed by someone else.

In particular, if propigateAdjustedCommits rebases the adjusted branch on top of master. That is called by sync. The assumption is that any changes in master have already been handled by updateAdjustedBranch. But, if another remote pushed a new master at just the right time, the adjusted branch could be rebased on top of a master that it doesn't incorporate, which is wrong.

Best fix seems to be to maintain a basis ref, that is not a branch, like refs/adjusted/master(unlocked). Copy master's ref to it when entering the view branch. Then, make all adjustments via the basis ref, and propigate back to refs/heads/master.

It's fine to overwrite changes that were pushed to master when propigating from the adjusted branch. Synced changes also go to synced/master so won't be lost. Pushes not made using git-annex sync of master are not really desired, just a possibility.

integration with view branches

Entering a view from an adjusted branch should probably carry the adjusting over into the creation/updating of the view branch.

Could go a step further, and implement view branches as another branch adjustment, albeit an extreme one. This might improve view branches. For example, it's not currently possible to update a view branch with changes fetched from a remote, and this could get us there.

This would need the reverse adjust to be able to change metadata, so that a commit that moved files in the view updates their metadata.

[WORKTREE: Wouldn't be able to integrate, unless view branches are changed into adjusted view worktrees.]

TODOs

  • Interface in webapp to enable adjustments.
  • Honor annex.thin when entering an adjusted branch. git checkout will make copies of the content of annexed files, so this would need to checkout the adjusted branch some other way. Maybe generalize so this more efficient checkout is available as a git-annex command?