Re: gc and repack ignore .git/*HEAD when checking reachability

From: Duy Nguyen <pclouds@gmail.com>
Date: Jul 10, 2016, 15:01 UTC
Message-ID: <CACsJy8AFYWsi1_f-Hw1odN3gmbfqayXuoBHo+K3Fh4XBwMA0Vw@mail.gmail.com>
In-Reply-To: <alpine.DEB.2.20.1607101602320.6426@virtualbox>

On Sun, Jul 10, 2016 at 4:16 PM, Johannes Schindelin <Johannes.Schindelin@gmx.de> wrote:

Show 11 quoted lines

>> No, the point is, refs subsystem needs to know which refs is per-repo,
>> which is per-worktree. So far the rules are  "everything under refs,
>> except a few known exceptions, is per-repo" and "everything directly
>> under $GIT_DIR is per-worktree", which work fine. But if you allow to
>> move per-worktree to "refs" freely, then the "known exceptions" will
>> have to be updated every time a new per-worktree ref appears. It'll be
>> easier to modify the first rule as "everything under refs, except some
>> legacy exceptions and refs/worktree, is per-repo".
>
> Given the substantial pain and suffering I have due to per-worktree
> reflogs (and those are *just* HEAD's reflogs!),

I apologize for that. I clearly did not see all the consequences of my simple-minded approach. Watch out for the shared config issue too if you start to rely on multiple worktrees. It may take even longer time to address than this one.

Show 9 quoted lines

> it appears to me that per-worktree refs would be a particularly poor feature.
>
> I agree that HEAD needs to be per-worktree, but already the fact that the
> HEADs of the worktrees, along with their reflogs, are *not* visible to
> all other worktrees causes substantial trouble.
>
> In my mind, a much more natural design would have been to map
> transparently each worktree's HEAD to refs/worktree/<name>/HEAD *and keep
> those refs and reflogs visible across all worktrees*.

We will be able to see refs from all worktrees if we choose to. There is no question about that. It's needed for full-repo operations, gc included. Whether we expose all worktree namespaces via refs/worktree (as a logical view, not storage one), remains to be seen. Michael's new ref API may be flexible enough that we could do without. I'm not sure yet.

Show 5 quoted lines

> With that design, I would not have irretrievably lost reflogs to auto-gc
> nor dozens of hours trying to figure out how to have *any* auto-gc succeed
> again (because I need to reduce the horrible slowness incurred by too many
> loose objects). My interactive rebases would also not have started to
> print the auto-gc error after every friggin' pick.

-- 
Duy