Re: git replace: should it check for object type, and can it replace merges?

Hi,

On Sat, Aug 3, 2013 at 5:13 PM, Philip Oakley [off-list ref] wrote:

A recent comment http://stackoverflow.com/a/18027030/717355 on a question I
asked two years ago about 'grafts' and 'replace' indicates that users think
that 'git replace' can't replace a merge commit. The documentation doesn't
have any examples and gives the naive impression that one should only
replace a simple commit with another simple commit.

I am sorry if the documentation gives this impression.
I'd like to fix it, but I am not sure what should be changed.
Should adding an example be enough? Or do you want it to state that explicitely?

Having looked at the code, I realised that anything can be replaced with
anything, which is perhaps not what was intended.

The documentation says in the "BUGS" section:

"And of course things may break if an object of one type is replaced
by an object of another type (for example a blob replaced by a
commit)."

So yes it is a know bug.

A simple thought is that
the replace should be like-for-like with regard to object type, though that
would not include replacing a sub-module for a tree (and vice versa).

Should 'git replace' check the object types to ensure they are sensible?

It would probably be a good idea to do that, yeah.
But I don't know much about submodules, so I can't say if replacing a
sub-module for a tree (and vice versa) should be allowed.
Or if there should be a --force-different-objects option for these
kinds of special cases.

Would it be reasonable to add examples to indicate the range of
replacements, and how to prepare alternative merge commits, or is that a
hostage to fortune?

Yeah, adding examples would be a good idea. I don't understand what do
you mean with "range of replacements", though.

I am not sure preparing alternative commits or merge commits should be
an important part of the examples.

There are many cases that could be interesting to different users:

- replacing a non merge commit with a merge commit (if someone forgot
to use --no-ff when merging for example)
- replacing a merge commit with a non merge commit (if a rebase should
have been done)
- and of course replacing a non merge commit with a non merge commit,
or a merge commit with a merge commit

So I think explaining how another commit can be created from existing
commits belongs to some other parts of the git documentation.
Perhaps there could be such examples in the git hash-object and git
filter-branch documentation and we could just point to them.

Best,
Christian.