On Wed, Dec 13, 2023 at 07:15:33AM -0800, Junio C Hamano wrote:
> Patrick Steinhardt <ps@pks.im> writes:
> 
> >> Me neither, but once you start thinking about getting rid of the
> >> need to use one-file-per-ref filesystem, being able to maintain all
> >> refs, including the pseudo refs, in one r/w store backend, becomes a
> >> very tempting goal.  From that point of view, I do not have problem
> >> with the idea to move _all_ pseudorefs to reftable.
> >
> > Yes, we're in agreement.
> >
> >> But I do have reservations on what Patrick, and the code he
> >> inherited from Han-Wen, calls "special refs" (which is not defined
> >> in the glossary at all), namely, refs.c:is_special_ref() and its
> >> callers.
> >
> > I do not want to add "special refs" to the glossary because ultimately
> > they should go away, with two exceptions: FETCH_HEAD and MERGE_HEAD.
> > Once we're there we can of course discuss whether we want to explicitly
> > point them out in the glossary and give them a special name.
> 
> OK, I somehow got a (wrong) impression that you are very close to
> the finish line.

You mean with the reftable backend? I indeed am quite close, I've just
finished the last prerequisite ("extensions.refFormat" and related
tooling) today. I will send that patch series upstream for review once
my patches that fix repo initialization with git-clone(1) land in the
"next" branch. The current state at [1] passes CI, even though there
will of course still be bugs which aren't covered by the test suite.

So once all prerequisites that are currently in flight plus the pending
"extensions.refFormat" series have landed I will send the reftable
backend implementation in for review. If things continue to go smoothly
I expect that this may happen at the end of January/start of February.

Anyway. This patch series here is in fact already sufficient to make
reftables work with those special refs. The only thing that we require
in this context is that refs are either exclusively routed through the
filesystem, or exclusively routed through the ref API. If that property
holds then things work just fine.

But still, I do want to clean up the remaining special refs regardless
of that, even though it is not a mandatory prerequisite. I find that the
current state is just plain confusing with all these special cases, and
I'd really love for it to be simplified. Also, I think there is benefit
in having those refs in reftables because it does allow for proper
atomic updates.

> If the intention is to leave many others still in
> the "special" category (for only the reasons of inertia), with a
> vision that some selected few must remain "special" with their own
> good reasons, then I am perfectly fine.

Okay.

Patrick

[1]: https://gitlab.com/gitlab-org/git/-/merge_requests/58