Re: [PATCH v2 1/5] replace: forbid replacing an object with one of a different type

From: Christian Couder <chriscool@tuxfamily.org>
Date: Aug 29, 2013, 06:29 UTC
Message-ID: <20130829.082936.1893875995125817253.chriscool@tuxfamily.org>
In-Reply-To: <xmqqa9k1a9wq.fsf@gitster.dls.corp.google.com>

From: Junio C Hamano <gitster@pobox.com>

Show 24 quoted lines

>
> Thomas Rast <trast@inf.ethz.ch> writes:
> 
>> Hrm, you're right, that's a flaw in my logic.  You could do the same in
>> all other cases too, e.g. replace a tree so that an entry is of a
>> different type and at the same time change the type of the object
>> itself.  You however have to carefully go through all objects that refer
>> to the one that was replaced, and fix the type in all of them.
>>
>> It still seems an extremely unsafe thing to do with trees...
>>  ...
>> Should we add a --force flag of some sort to allow the user to do this,
>> while keeping the normal safety checks?
> 
> As long as we do not forbid such an unusual replacement on the
> reading side, we won't break people who are more inventive than we
> are (I am not convinced that we know people's workflow well enough
> to definitively say that no sane workflow, which benefits from being
> able to replace an object with another from a different type,
> exists).
> 
> Preventing "git replace" wrapper from creating such a replacement by
> default will make it harder to do and may reduce mistakes, without
> breaking them too much, I think.

I agree. It is always possible to create replacement refs using "git update-ref" if one really wants to.

What about using the following in the commit message or in the documentation:

------------------------------------------

DISCUSSION

If one object is replaced with one of a different type, the only way to keep the history valid is to also replace all the other objects that point to the replaced object. That's because:

* 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.

But if all the objects that point to an object, called O, are to be replaced, then in most cases object O probably doesn't need to be replaced. It's probably sufficient to create the new object, called O2, that would replace object O and to replace all the objects pointing to object O with objects pointing to O2.

The only case where someone might really want to replace object 0, with an object O2 of a different type, and all the objects pointing to it, is if it's really important, perhaps for external reasons, to have object O's SHA1 point to O2.

And anyway, if one really wants to do that, it can still be done using "git update-ref".

------------------------------------------

Thanks, Christian.