git object sharing is suboptimal in Launchpad

IMO this is only a problem because we haven't yet taken a project-wide decision to bless the importer branches as the default repositories for their corresponding source packages (i.e. lp:ubuntu/+source/PACKAGE). Once that happens, object sharing will automatically be set up for newly-pushed repositories.

On Tue, Jul 25, 2017 at 10:43:19AM -0000, Colin Watson wrote:
> IMO this is only a problem because we haven't yet taken a project-wide
> decision to bless the importer branches as the default repositories for
> their corresponding source packages (i.e. lp:ubuntu/+source/PACKAGE).
> Once that happens, object sharing will automatically be set up for
> newly-pushed repositories.

FTR, we probably don't want to do that until the importer team declares
"commit hash stability" (IOW, a branch fast-forwarding guarantee). So
this may merely become an ordering issue. We may not want to do this
until we've seen wider use to gain confidence, and may not want wider
use until we've got object sharing sorted :)

An example is how we're handling empty directories in source packages
(bug 1687057). We discovered that problem only after some level of
increased use, and the current plan (implemented in the importer) is to
put them into the git commit DAG even though the git porcelain doesn't
yet support them. This makes all previously imported commits of sources
with empty directories "wrong", which makes us want to re-import and
break fast-forwarding. Ideally we'd avoid declaring commit hash
stability until upstream have accepted empty directory support in git
porcelain (yet to be written).

There may be other issues like this one that we have yet to discover.

At some point we'll have to draw a line I expect, and accept any
historical errors in the importer in already import commits after that
point. But it'd be nice to get wider testing before we do, so I'm not
sure where that sits with the lack of git object sharing with the
repositories as they are at the moment.

Yes, I see that point.

In theory it would be possible to designate a repository as the default baseline for object sharing even though it's not the default for name lookup, but it would be quite a lot of extra configuration complexity and so I think it's probably not a good idea.

We've discussed having an explicit fork action in the past, which would give us explicit knowledge of the repository to object-share with rather than having to guess, and would let us tell users what URL to push to rather than relying on them knowing how Launchpad's Git URL namespace is laid out. I'm not opposed to that, but it would be an extra chunk of work.

Could we perhaps put the importer branches in their "official" places, but make it clear that they are experimental and non-fast-forwarding until we declare hash stability at some future 1.0 date? I'm not sure I'm keen on this, but it may be the least worst option.

@racb: do you have a spec of the default aliases (even pseudocode, or what is necessary where)? I can work on implementing it, if so.