Re: [PATCH/RFC 4/6] transport: add refspec list parameters to functions

[David Turner <dturner@twopensource.com>]

On Wed, 2016-04-20 at 16:57 -0400, Jeff King wrote:
> On Wed, Apr 20, 2016 at 04:46:55PM -0400, David Turner wrote:
> 
> > As you note, it appears that git-daemon does sort-of have support
> > for
> > extra args -- see parse_host_arg.  So it wouldn't be hard to add
> > something here. Unfortunately, current versions of git die on
> > unknown
> > args.  So this change would not be backwards-compatible.  We could
> > put
> > a decider on it so that clients would only try it when explicitly
> > enabled.  Or we could have clients try it with, and in the event of
> > an
> > error, retry without.  Neither is ideal, but both are possible.
> 
> Right. This ends up being the same difficulty that the v2 protocol
> encountered; how do you figure out what you can speak without
> resorting
> to expensive fallbacks, when do you flip the switch, do you remember
> the
> protocol you used last time with this server, etc.

Right.

[moved]
> > If I read this code correctly, git-over-ssh will pass through
> > arbitrary
> > arguments.  So this should be trivial.
> 
> It does if you are ssh-ing to a real shell-level account on the
> server,
> but if you are using git-shell or some other wrapper to restrict
> clients
> from running arbitrary commands, it will likely reject it.

Oh, I see how I was mis-reading shell.c.  Oops.
[/moved]


> Which isn't to say it's necessarily a bad thing. Maybe the path
> forward
> instead of v2 is to shoe-horn this data into the pre-protocol
> conversation, and go from there. The protocol accepts that "somehow"
> it
> got some extra data from the transport layer, and acts on its
> uniformly.

OK, so it seems like only HTTP (and non-git-shell-git://) allow backwar
ds-compatible optional pre-protocol messages.  So we don't have good
options; we only have bad ones.  We have to decide which particular
kind of badness we're willing to accept, and to what degree we care
about extensibility.  As I see it, the badness options are (in no
particular order):

1. Nothing changes.
2. HTTP grows more extensions; other protocols stagnate.
3. HTTP grows extensions; other protocols get extensions but:
   a. only use them on explicit client configuration or
   b. try/fail/remember per-remote
4. A backwards-incompatible protocol v2 is introduced, which
   hits alternate endpoints (with the same a/b as above).  This is
   different from 3. in that protocol v2 is explicitly designed around
   a capabilities negotiation phase rather than unilateral client-side
   decisions.
5. Think of another way to make fetch performant with many refs, and 	
    defer the extension decision.