The case for two trees in a commit ("How to make rebase less modal")

$ git hash-object --stdin -w -t commit <<EOF tree c70b4a33a0089f15eb3b38092832388d75293e86 parent 105d5b91138ced892765a84e771a061ede8d63b8 author Stefan Beller <sbeller@google.com> 1519859216 -0800 committer Stefan Beller <sbeller@google.com> 1519859216 -0800 tree 5495266479afc9a4bd9560e9feac465ed43fa63a test commit EOF 19abfc3bf1c5d782045acf23abdf7eed81e16669 $ git fsck |grep 19abfc3bf1c5d782045acf23abdf7eed81e16669 $

So it is technically possible to create a commit with two tree entries and fsck is not complaining.

But why would I want to do that?

There are multiple abstraction levels in Git, I think of them as follows:
* data structures / object model
* plumbing
* porcelain commands to manipulate the repo "at small scale", e.g.
create a commit/tag
* porcelain to modify the repo "at larger scale", such as rebase,
cherrypicking, reverting
  involving more than 1 commit.

These large scale operations involving multiple commits however are all modal in its nature. Before doing anything else, you have to finish or abort the rebase or you need expert knowledge how to go otherwise.

During the rebase there might be a hard to resolve conflict, which you may not want to resolve right now, but defer to later. Deferring a conflict is currently impossible, because precisely one tree is recorded.

If we had multiple trees possible in a commit, then all these large scale operations would stop being modal and you could just record the unresolved merge conflict instead; to come back later and fix it up later.

I'd be advocating for having multiple trees in a commit possible locally; it might be a bad idea to publish such trees.

Opinions or other use cases?

Thanks, Stefan

On 28/02/18 23:30, Stefan Beller wrote:

Hmm, it's a while since I looked at that code, but I don't think you have a commit with two trees - the second 'tree <sha1>' line is just part of the commit message, isn't it?

ATB, Ramsay Jones

On 03/01, Ramsay Jones wrote:

Actually it doesn't look like it. The commit msg doesn't start till after an empty newline so that commit has an empty commit msg. Here's one which you can see the msg when passed to show:

git hash-object --stdin -w -t commit <<EOF tree 76d269b57d3c4283922216f84a2850e99f561ccc parent fa0624f79f9d5765d09598b003124b3cf0b9acdb author Brandon Williams <bmwill@google.com> 1519859216 -0800 committer Brandon Williams <bmwill@google.com> 1519859216 -0800 tree 76d269b57d3c4283922216f84a2850e99f561ccc

This is a test commit with multiple trees EOF

Of course the extra tree is ignored, but fsck doesn't complain and show happily shows what it knows about.

-- Brandon Williams

On Wed, Feb 28, 2018 at 03:30:27PM -0800, Stefan Beller wrote:

What benefit does it have over adding a new header "unresolved-tree" or similar? I do not think you are getting any backwards compatibility here. For instance, "prune" will not traverse it with existing versions of git, nor "pack-objects" include it in a pack (I didn't actually test it, so I could be wrong; but those are all based around parse_commit, which should look at only the first tree).

-Peff

On Wed, Feb 28, 2018 at 3:30 PM, Stefan Beller <sbeller@google.com> wrote:

How does this let you defer a conflict? A future commit which modified blobs in that tree wouldn't know what version of the trees/blobs to actually use? Clearly future commits could record their own trees, but how would they generate the "correct" tree?

Maybe I am missing something here?

Thanks, Jake

Jacob Keller <jacob.keller@gmail.com> writes:

If you write four trees out of each stage in the index and record them, you could in theory have a new command that reads them and recreate the conflicted index. Oh, and then you would need the fifth tree that records what the working-tree files (with conflict markers) looked like, in order to reproduce the state seen by the person who ran "git merge", attempted to resolve and gave up halfway in the middle.

As a local operation, extending "git stash" somehow so that it can stash even in a conflicted working tree may be a better approach, and it does not need cruft headers in commit objects, I would think.

Stefan Beller <sbeller@google.com> writes:

The second one is merely a random unauthorized header that is not interpreted in any way by Git. It is merely being confusing by starting with "tree " and having 40-hex after it, but the 40-hex does not get interpreted as an object name, and does not participate in reachability computation (i.e. packing, pruning and fsck).

There is not much difference between that and a line of trailer in the commit log message (other than this one is less discoverable).

Hi,

Stefan Beller wrote:

As others mentioned, this is essentially a fancy way to experiment with adding a new header (with the same name as an existing header) to a commit. It is kind of a scary thing to do because anyone trying to parse commits, including old versions of git, is likely to get confused by the multiple trees. It doesn't affect the reachability calculation in the way that it should so this ends up being something that should be straightforward to do with a message in the commit body instead.

To affect reachability, you could use multiple parent lines instead. You'd need synthetic commits to hang the trees on. This is similar to how "git stash" stores the index state.

In other words, I think what you are trying to do is feasible, but not in the exact way you described.

[...]

Junio mentions you'd want to record:
 - stages of the index, to re-create a conflicted index
 - working tree files, with conflict markers

In addition you may also want to record:
 - state (todo list) from .git/rebase-merge, to allow picking up where
   you left off in such a larger operation
 - similar state for other commands --- e.g. MERGE_MSG

Recording this work-in-progress state is in the spirit of "git stash" does. People also sometimes like to record their state in progress with a "wip commit" at the tip of a branch. Both of those workflows would benefit from something like this, I'd think.

So I kind of like this. Maybe a "git save-wip" command that is like "git stash" but records state to the current branch? And similarly improving "git stash" to record the same richer state.

And in the spirit of "git stash" I think it is possible without even modifying the commit object format.

[...]

I think such "WIP state" may also be useful for publishing, to allow collaborating on a thorny rebase or merge.

Thanks and hope that helps, Jonathan