Re: [PATCH v2 40/43] refs: allow ref backend to be set for clone

On Tue, 2015-10-06 at 14:09 -0400, David Turner wrote:

On Mon, 2015-10-05 at 21:58 -0400, Jeff King wrote:

quoted

On Mon, Oct 05, 2015 at 09:29:37PM -0400, David Turner wrote:

quoted

quoted

Therefore, I don't think this can be merged without a bump to
core.repositoryformatversion. Such a bump will tell well-
behaved older
Git clients keep their hands off the repository. (Of course
repositories
that use the files backend can continue using
core.repositoryformatversion 0.)

I thought Peff proposed a way to do such a bump, including a
way to
extend repositories one by one with new features. But that was
something
that we were chatting about off-list.

I haven't reviewed the actual code in this patch yet but I
wanted to get
the above comment on your radar.

Michael

I'll fix this to upgrade to v=1 when the lmdb refs backend is in
use,
and to give sensible error messages in a v1 repo if built without
LMDB.

I think the relevant series is:

  http://article.gmane.org/gmane.comp.version-control.git/272447

It did not seem too controversial, but it mostly got dropped amidst
the
release, and I haven't reposted it yet.

That patch will work perfectly for this use case.  I'll add it to my
series when I reroll, and set an extension.

This is something I'm working on right now for libgit2 as well; not
lmdb specifically but allowing user-provided backends, which would
allow built-in ones as well.

Did we ever decide on the format for these extensions? The series
mentioned above has a couple of examples, but doesn't have any testing
for backend stuff. Do we have a concrete proposal for this? I was going
to go for something like

extensions.refbackend = "lmdb"
extensions.odbbackend = "psql"

and have backends register themselves by the "lmdb", "psql" or whatever
format, but if we have already decided something else which I missed,
I'd swap over to that.

Cheers,
   cmn