On Thu, May 30, 2019 at 9:19 PM Bryan Turner [off-list ref] wrote:

I was looking through the commit history in a repository I work in and
I found a place where someone had created a merge, but somewhere
between "git merge" and "git commit" the fact that it was a merge was
"lost". Instead they ended up with a really big commit that applied
all the changes from the merged-in branch.

A really easy way to reproduce this is:
git merge master #Assume this has conflicts, or use --no-commit
git checkout -b some-new-branch

When the checkout runs, MERGE_HEAD et al are deleted without any sort
of warning, but the uncommitted changes are not lost. If a user then
runs "git commit", and doesn't notice that there's no helpful "It
looks like you may be committing a merge", they'll create a new,
non-merge commit that essentially reapplies all the changes they
merged in.

I'm pretty familiar with Git and I make this mistake at least a few
times a year. So far I've always caught it during the commit, though.
Unfortunately, in this case, the bad "merge" wasn't noticed before it
made its way to master, so now it's there for good.

I'm not sure what there is to do about this. It's clear it's a
long-standing behavior. One approach might be to introduce a warning
when changing branches deletes MERGE_*. A different one might be to
fail to change branches without something like --force. I'm not sure
either is _better_ than the current behavior, but they're certainly
_clearer_. That said, perhaps this behavior is something someone
relies on.

Best regards,
Bryan Turner

Discussed in detail recently starting at
https://public-inbox.org/git/CACsJy8Axa5WsLSjiscjnxVK6jQHkfs-gH959=YtUvQkWriAk5w@mail.gmail.com/

resulting in
https://public-inbox.org/git/20190329103919.15642-8-pclouds@gmail.com/
and
https://public-inbox.org/git/20190329103919.15642-24-pclouds@gmail.com/

I think we should still do the follow up for checkout, as mentioned at
https://public-inbox.org/git/CABPp-BHX1gRhTdurAwrPg60Hk-OuhbrEN=4zatx4OOUo-DkQvw@mail.gmail.com/

It's good to get extra feedback that this isn't just theoretical but
is causing people actual problems.  Do you want to take the time to
make the change I suggested in the last email above and propose it to
the list?  I think the main thing needed is just a good commit message
and getting feedback and thoughts from others; your description above
was well written and I'm busy on other things right now, so if you'd
like to tackle it, I'd appreciate it.  If not, I will hopefully
remember to get back to it eventually.

Hope that helps,
Elijah