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

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

[Philip Oakley]

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.

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

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?


Philip Oakley

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

[Thomas Rast]

"Philip Oakley" <philipoakley@iee.org> writes:

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

Off hand I cannot come up with any case where you can replace one object
with one of a different type, keeping the history valid.

To back that up:

* Refs can be "replaced" simply by changing the ref itself.

* Annotated tags contain the type of the tagged object.

* The tree/parent lines in commits must be a tree and commits, resp.

* The object types referred to by trees are specified in the 'mode'
  field:
    100644 and 100755    blob
    160000               commit
    040000               tree
  (these are the only valid modes)

* Blobs don't point at anything.

So yes:

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

I think that would be a very good thing to check.

-- 
Thomas Rast
trast@{inf,student}.ethz.ch

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

[Christian Couder]

Hi,

On Sat, Aug 3, 2013 at 5:13 PM, Philip Oakley <philipoakley@iee.org> 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.

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

[Philip Oakley]

From: "Christian Couder" <christian.couder@gmail.com>
> Hi,
> On Sat, Aug 3, 2013 at 5:13 PM, Philip Oakley <philipoakley@iee.org> 
> 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?
>
I did a quick markup which is shown below (an export of the commit from 
the gitk viewer)

>> 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)."

I hadn't seen that part, being 'hidden' at the end of the paragraph 
(that is, it's easy to miss;-). If my update was acceptable then that 
sentence could probably be deleted (though it may require the checks to 
actually be in the code, and not just a "must" imperative in my 
documentation update).

>
> 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.

An extra bit of thought made me realise that while a sub-module is 
represented as a special symbolic commit, it is still just an element of 
a tree object, so would still be a tree <-> tree replacement, so doesn't 
break the rule.

>
>> 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.

There were in two parts: 1) creating a replacement merge commit, and 2) 
creating a replacement tree,  possibly with a sub-module in it.

>
> 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.

Yes, I just haven't looked yet. I think there are some examples in the 
plumbing command man pages. They'd just need referencing.

> 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.
> --

My quick markup, based on a local branch.
---8<---
commit c12c03462f8c65a593e702896b461f1c63d67ec5
Author: Philip Oakley <philipoakley@iee.org>
Date:   Sat Aug 3 20:20:05 2013 +0100

    Doc: 'replace' the same object type, and mention merge commits

    Signed-off-by: Philip Oakley <philipoakley@iee.org>

diff --git a/Documentation/git-replace.txt 
b/Documentation/git-replace.txt
index e0b4057..2ab451c 100644
--- a/Documentation/git-replace.txt
+++ b/Documentation/git-replace.txt
@@ -20,6 +20,10 @@ The name of the 'replace' reference is the SHA-1 of 
the object that is
 replaced. The content of the 'replace' reference is the SHA-1 of the
 replacement object.

+The type of the replacement object must be the same as that of the
+object it replaces. Merge commits can be replaced by non-merge commits
+and vice versa.
+
 Unless `-f` is given, the 'replace' reference must not yet exist.

 Replacement references will be used by default by all Git commands
--->8---

Philip 

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

[Junio C Hamano]

Christian Couder <christian.couder@gmail.com> writes:

> Hi,
>
> On Sat, Aug 3, 2013 at 5:13 PM, Philip Oakley <philipoakley@iee.org> 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.

Is that even a BUG?  The users are pretty much asking for it if they
place an object name of a random wrong object themselves.

I agree that a hand-holding wrapper "git replace" could help them to
avoid mistakes, though.