Is GIT_DEFAULT_HASH flawed?

From: Felipe Contreras <felipe.contreras@gmail.com>
Date: May 2, 2023, 23:46 UTC
Message-ID: <6451a0ba5c3fb_200ae2945b@chronos.notmuch>
In-Reply-To: <20230427054343.GE982277@coredump.intra.peff.net>

Hi,

Changing the subject as this message seems like a different topic.

Jeff King wrote:

Show 28 quoted lines

> On Wed, Apr 26, 2023 at 02:33:30PM -0700, Junio C Hamano wrote:
> > "brian m. carlson" <sandals@crustytoothpaste.net> writes:
> > 
> > >  `GIT_DEFAULT_HASH`::
> > >  	If this variable is set, the default hash algorithm for new
> > >  	repositories will be set to this value. This value is currently
> > > +	ignored when cloning if the remote value can be definitively
> > > +	determined; the setting of the remote repository is used
> > > +	instead. The value is honored if the remote repository's
> > > +	algorithm cannot be determined, such as some cases when
> > > +	the remote repository is empty. The default is "sha1".
> > > +	THIS VARIABLE IS EXPERIMENTAL! See `--object-format`
> > > +	in linkgit:git-init[1].
> > 
> > We'd need to evantually cover all the transports (and non-transport
> > like the "--local" optimization) so that the object-format and other
> > choices are communicated from the origin to a new clone anyway, so
> > this extra complexity "until X is fixed, it behaves this way, but
> > otherwise the variable is read in the meantime" may be a disservice
> > to the end users, even though it may make it easier in the shorter
> > term for maintainers of programs that rely on the buggy "git clone"
> > that partially honored this environment variable.
> > 
> > In short, I am still not convinced that the above is a good design
> > choice in the longer term.
> 
> I also think it is working against the backwards-compatible design of
> the hash function transition.

To be honest this whole approach seems to be completely flawed to me and against the whole design of git in the first place.

In a recent email Linus Torvalds explained why object ids were calculated based {type, size, data} [1], and he explained very clearly that two objects with exactly the same data are not supposed to have the same id if the type is different.

If even the tiniest change such as adding a period to a commit messange changes the object id (and thus semantically makes it a different object), then it makes sense that changing the type of an object also changes the object id (and thus it's also a different object).

And because the id of the parent is included in the content of every commit, the top-level id ensures the integrity of the whole graph.

But then comes this notion that the hash algorithm is a property of the repository, and not part of the object storage, which means changing the whole hash algorithm of a repository is considered less of a change than adding a period to the commit message, worse: not a change at all.

I am reminded of the warning Sam Smith gave to the Git project [2] which seemed to be unheard, but the notion of cryptographic algorithm agility makes complete sense to me.

In my view one repository should be able to have part SHA-1 history, part SHA3-256 history, and part BLAKE2b history.

Changing the hash algorithm of one commit should change the object id of that commit, and thus make it semantically a different commit.

In other words: an object of type "blob" should never be confused with an object of type "blob:sha-256", even if the content is exactly the same.

The fact that apparently it's so easy to clone a repository with the wrong hash algorithm should give developers pause, as it means the whole point of using cryptographic hash algorithms to ensure the integrity of the commit history is completely gone.

I have not been following the SHA-1 -> OID discussions, but I distinctively recall Linus Torvalds mentioning that the choice of using SHA-1 wasn't even for security purposes, it was to ensure integrity. When I do a `git fetch` as long as the new commits have the same SHA-1 as parent as the SHA-1s I have in my repository I can be relatively certain the repository has not been tampered with. Which means that if I do a `git fetch` that suddenly brings SHA-256 commits, some of them must have SHA-1 parents that match the ones I currently have. Otherwise how do I know it's the same history?

Maybe that's one of the reasons people don't seem particularly eager to move away from SHA-1:

Better the SHA-1 you know, than the SHA-256 you don't.

Cheers.

[1] https://lore.kernel.org/git/CAHk-=wjr-CMLX2Jo2++rwcv0VNr+HmZqXEVXNsJGiPRUwNxzBQ@mail.gmail.com/ [2] https://lore.kernel.org/git/D433038A-2643-4F63-8677-CA8AB6904AE1@samuelsmith.org/

-- 
Felipe Contreras