Re: [PATCH v3 2/2] builtin/pack-objects.c: introduce `pack.recentObjectsHook`

From: Taylor Blau <me@ttaylorr.com>
Date: May 12, 2023, 21:36 UTC
Message-ID: <ZF6xT34/Zq5pT9fG@nand.local>
In-Reply-To: <20230512212456.GA2495860@coredump.intra.peff.net>

On Fri, May 12, 2023 at 05:24:56PM -0400, Jeff King wrote:

Show 28 quoted lines

> > This patch introduces a new configuration, `pack.recentObjectsHook`
> > which allows the caller to specify a program (or set of programs) whose
> > output is treated as a set of objects to treat as recent, regardless of
> > their true age.
>
> I was going to complain about putting this in the "pack" section,
> because I thought by touching reachable.c, we'd also affect git-prune.
> But I don't think we do, because it does its own direct mtime check on
> the loose objects.
>
> But I'm not sure that's the right behavior.
>
> It feels like even before your patch, this is a huge gap in our
> object-retention strategy.  During repacking, we try to avoid dropping
> objects which are reachable from recent-but-unreachable things we're
> keeping (since otherwise it effectively corrupts those recent objects,
> making them less valuable to keep). But git-prune will happily drop them
> anyway!
>
> And I think the same thing would apply to your hook. If the hook says
> "object XYZ is precious even if unreachable, keep it", then git-prune
> ignoring that seems like it would be a source of errors.
>
> I suspect both could be fixed by having git-prune trigger the same
> add_unseen_recent_objects_to_traversal() call either as part of
> the perform_reachability_traversal() walk, or maybe in its own walk (I
> think maybe it has to be its own because the second walk should avoid
> complaining about missing objects).

I might be missing something, but I think we already (kind of) do the right thing here.

AFAICT, the path is:

  - cmd_prune()
  - for_each_loose_file_in_objdir()
  - prune_object() (as a callback to the above)
  - is_object_reachable()
  - perform_reachability_traversal()
  - mark_reachable_objects()
  - add_unseen_recent_objects_to_traversal()

That only happens when `mark_recent != 0`, though.

Thanks, Taylor