On Thu, Jan 13, 2022 at 07:24:19PM +0100, Han-Wen Nienhuys wrote:
> On Fri, Jan 7, 2022 at 11:17 PM Junio C Hamano <gitster@pobox.com> wrote:
> > > this is a resend of version 1 of this patch series to hopefully entice
> > > some reviews. The only change is that v2 is rebased onto the current
> > > main branch at commit e83ba647f7 (The seventh batch, 2022-01-05). The
> > > following was from the orignial cover letter:
> >
> > I'll add Ævar, who has been making a lot of changes to the refs
> > subsystem, and Han-Wen, whose work to add a new ref backend may need
> > to interact with this change, as possible stake-holders to the CC list.
> 
> Thanks for the consideration, Jun. As the hook is called from refs.c,
> it should not make a difference for reftable.
> 
> I looked over the patches. I didn't look at the bottom change to
> packed/loose refs as I'm not an expert.
> 
> The individual transaction updates already support their own flags, so
> this change generates some confusion. Consider:
> 
> int refs_delete_ref(struct ref_store *refs, const char *msg,
>     const char *refname,
>     const struct object_id *old_oid,
>     unsigned int flags)
> 
> how would one delete a ref skipping the transaction hook? It will be
> easy for someone to pass the SKIP_TRANSACTION_HOOK to
> refs_delete_ref(), with surprising results.
> 
> It might make sense to not introduce a new flag namespace, but use
> update->flags instead. You'd have to add your new flag after
> REF_SKIP_REFNAME_VERIFICATION.
> Bonus is that you can unittest the new flag using the existing
> ref-store helper without extra work. (check that a transaction with &
> without the flag works as expected.)
> 
> other than that, looks OK to me.

Thanks for your feedback!

In fact the first version I had locally did exactly that. But I found
the end version result harder to reason about, most importantly because
it's not a 100% clear to the reader whether all callsites which delete
refs in the packed-backend via the files-backend are adapted or whether
any of the callsites was missing. By having the flag in a central place
it's immediately clear that the hook won't be run at all, which is
exactly what we want here.

Patrick