On Fri, Aug 03 2018, Jeff King wrote:
> On Fri, Aug 03, 2018 at 02:56:11PM +0200, Ævar Arnfjörð Bjarmason wrote:
>
>> > Any Git commands you run should therefore find objects from either
>> > location, but any writes would go to the quarantine (most notably, Git's
>> > own index-pack/unpack-objects processes, which is the point of the
>> > quarantine in the first place).
>>
>> To add to this, one interesting thing that you can do with hooks because
>> of this quarantine is to answer certain questions about the push that
>> were prohibitively expensive before it existed, but there's no explicit
>> documentation for this.
>>
>> E.g. for a hook that wants to ban big blobs in the repo, but wants to
>> allow all existing blobs (you don't want to block e.g. a revert of a
>> commit that removed it from the checkout), you can juggle these two env
>> variables and hide the "main" object dir from the hook for some
>> operations, so e.g. if a blob lookup succeeds in the alternate
>> quarantine dir, but not the main object dir, you know it's new.
>
> I'd be a bit careful with that, though, as the definition of "new" is
> vague there.
>
> For example, completing a thin pack may mean that the receiver creates a
> copy of a base object found in the main repo. That object isn't new as
> part of the push, nor was it even sent on the wire, but it will appear
> in the quarantine directory. But only sometimes, depending on whether we
> kept the sender's pack or exploded it to loose objects.
Right, I mean:
is_new = !in_quarantine() && in_main()
Or:
is_new = !in_main()
Should work, in the latter case if the object really is missing from the
quarnatine too, other fsck bits will stop the push.
But as you point out:
is_new = in_quarantine()
Cannot be relied upon, although it'll be true most of the time.
Perhaps I'm missing some edge case above, but I wanted to reword it to
make sure I understood it correctly (and perhaps you have a correction).
