On Tue, May 27, 2025 at 02:45:43PM -0500, Justin Tobler wrote:

On 25/05/27 09:50AM, Junio C Hamano wrote:

quoted

Patrick Steinhardt [off-list ref] writes:

quoted

I think the only outstanding discussion is whether to name things
`odb_alternate` or `odb_source` [1]. In case others agree that
`odb_source` is a better name I'm happy to revise, but if not I'd rather
keep it as-is.

The model in which the term "alternates" was born is "A repository
has its own object directory, the primary one, and in addition it
can borrow from zero or more alternate object directories that are
used by other repositories".  The presence of the primary makes the
word "alternate" meaningful.

Is the model now "A repository has one object store, which consists
of one or more X, all of which are equals"?  If there is no primary
that is more special than others, then calling X an "alternate" may
indeed sound funny, although (1) I do not find it terribly confusing
and (2) I do not find "source" much better, either.

My understanding is that the object store still has a primary X and zero
or more alternative X. The idea is that eventually, with pluggable ODBs,
X can be a different backend/provider instead of just being "files". If
this is the case, calling X an "alternate" would mean we have a primary
"alternate" and potentially a set of "alternate" alternates.

This sounds a bit odd and doesn't quite match what I would intuitively
expect. But, I also don't find it super confusing either.

Yeah, I understand that confusion indeed. I don't think that the other
proposals we've got are a lot better, either:

  - `odb_backend` was shot down because it may cause the association
    that one object database has one backend. But backends are per
    alternate, so there's a mismatch in expectations.

  - `odb_source` is better, but we now have the problem that we use
    "alternate" interchangably in most cases where we also use
    `odb_source`. This will likely lead to somewhat awkward interfaces.

The problem with `odb_source` might eventually go away once we clearly
distinguish the "alternates" concept from the low-level mechanism to
access objects. But I'm just not certain at all whether it won't cause
more confusion when in most cases "alternates" and "sources" can be used
somewhat interchangably.

I dunno. The more I think about this the more I start to like the
`odb_source` name.

quoted

The names we use to call the collection and the underlying
implementations of the collection in the reference world
unfortunately does not quite help to guide us, as we do not take two
implementations and compose into one unified view, which is what we
are doing in the object store.  Hmmm...

Similar to references, I still think of a pluggable ODB as a "backend".
The main difference being that with references there is only a single
backend active ("file" or "reftables") at a time, while for the object
store there could be multiple.

Yeah, that's basically the major source of confusion I want to avoid
with the "backend" terminology. I really do want to make it explicit
that there is not only one backend, but multiple ones.

Patrick