Disappearing change on pull rebase

Hi all, I've got an issue with some operations. It seems like git eats one of my commits (it's still in reflog, but in normal tree, it's unavailable).

What I did is:

checkout -b feature/.... (edit files and commit) checkout master merge --no-ff --no-commit feature/... (edit some files, change versions, changelog) commit

Now I've got the change committed in the branch and some more changes on the merge commit. So before pushing to the main repo, I'd like to check if any other changes are there:

pull --rebase

Now my merge commit disappears completely along with the changes without any warning. I get the branch commits duplicated on top of master and the branch stays as it was. That looks like a data loss bug to me since I can only recover a committed change from reflog and there are no warnings before that change goes away (using 1.7.4.1). Actually no changes were done in upstream in the meantime, so the rebase was not even needed.

Regards, Stanisław Pitucha Cloud Services Hewlett Packard

10.11.2011, в 16:15, Pitucha, Stanislaw Izaak:

That is definitely not a bug. "git pull --rebase" is (almost?) equal to "git fetch ; git rebase origin/master" When you perform a rebase, at first your HEAD is rolled back to the commit before your changes, then it is fast-forwared to the remote HEAD (in your case, no fast-forward was made, because there were no remote changes); then your commits are applied one by one.

Of couse, when your commits are applied, they are applied like patches. That mean, that they are different from the original commits (at least, by the commit time). That causes the duplication.

And the merge commit "dissapeared", because it contained no changes, so the patch was empty, and there was nothing to reapply. If the merge commit contained some changes, and it really was not applied during rebase, it is a bug, but more details will be needed, I think.

If you want to preserve your branch history, you should do "pull" without "rebase".

Kirill.

As mentioned in the original mail - the merge commit did have changes. Here's the log of reproducing it. The line containing "2" in changelog is gone from master after pull --rebase.

$ git init fake_upstream
Initialized empty Git repository in /tmp/fake_upstream/.git/
$ cd fake_upstream/ 
fake_upstream$ echo some_content > test
fake_upstream$ echo 1 > changelog
fake_upstream$ git add test changelog 
fake_upstream$ git commit -m 'first commit'
[master (root-commit) b1b683d] first commit
 2 files changed, 2 insertions(+), 0 deletions(-)
 create mode 100644 changelog
 create mode 100644 test
fake_upstream$ cd ..

$ git clone /tmp/fake_upstream/ disappearing_commit
Cloning into disappearing_commit...
done.
$ cd disappearing_commit/
disappearing_commit$ git checkout -b some-branch
Switched to a new branch 'some-branch'
disappearing_commit$ echo blah >> test
disappearing_commit$ git commit test -m 'branch modification'
[some-branch 528f166] branch modification
 1 files changed, 1 insertions(+), 0 deletions(-)
disappearing_commit$ git checkout master
Switched to branch 'master'

disappearing_commit$ git merge --no-ff --no-commit some-branch Automatic merge went well; stopped before committing as requested disappearing_commit$ echo 2 >> changelog disappearing_commit$ git add changelog disappearing_commit$ git commit [master e41e4c9] Merge branch 'some-branch'

disappearing_commit$ git show
commit e41e4c963a72528e3b2b5945ca419e19cdba1e4b
Merge: b1b683d 528f166
Author: Stanislaw Pitucha <stanislaw.pitucha@hp.com>
Date:   Thu Nov 10 13:30:04 2011 +0000

    Merge branch 'some-branch'

diff --cc changelog index d00491f,d00491f..1191247

Am 11/10/2011 14:35, schrieb Pitucha, Stanislaw Izaak:

This is by design. Rebase does not rebase merge commits because it is assumed that merge commits only do what their name implies - to merge branches of a forked history. As such, they do not introduce their own changes. Follow this rule, i.e., make your change in a separate non-merge commit, and you are fine.

-- Hannes

Doesn't this create a problem if you pull, resolve a conflit but do NOT push, then pull --rebase some more commits later on? As I understand it, the conflict resolution commit will be a merge commit and will be thrown away by the git pull --rebase.

Philippe

Am 11/11/2011 10:50, schrieb Philippe Vaucher:

You are correct, but it is not a problem: During the rebase, the same conflicts will arise as during the merge, and you will be forced to resolve them before you can complete the rebase. Therefore, nothing will be lost.

-- Hannes

Doing the same conflict resolution twice is kinda irritating... but ok, fair enough. I guess when you resolve a conflict you're supposed to push to avoid this :)

Philippe

Johannes Sixt <j.sixt@viscovery.net> writes:

While that may be technically correct, I wonder if we can be a bit more helpful to the users in such a case (upfront I admit that I have a strong suspicion that this is a hard problem in general). One typical use of "rebase" is to linearize a history that has unfortunate merges that did not have to be there, so refusing "git rebase A..B" when there is a merge in "git rev-list --merges A..B" is not a solution. But would it help if we warned about the merges when we find that there is something _interesting_ going on in them, e.g. an evil merge that adds material that did not exist in any of the parents [*1*]? The warning message may diagnose "you asked me to linearize the history by picking commits on the non-merge segments and replaying them, but there may be changes made in this merge commit, and it does this interesting thing: $(git show -c $that_merge_commit)" and may further suggest "if you do not want to linearize but just transplant the history, perhaps you want to run the command with the '-m' option?".

[Footnote]

*1* This is a hard problem and not just the matter of looking at "show -c" output. A "-s ours" merge would appear empty in "show -c" but it _is_ an interesting event that makes the result of linearizing non-merge segments vastly different from the original. Also material that did not exist in any of the parents is not necessarily evil (e.g. the side branch may have added one parameter to a function and updated its call sites, while our branch may have added a different parameter to the same function. The update to the call sites in the merge result should pass two more parameters from the common ancestor, and different from either of the parents).