Re: Pathological performance with git remote rename and many tracking refs

From: Junio C Hamano <gitster@pobox.com>
Date: Apr 15, 2022, 17:25 UTC
Message-ID: <xmqqfsmedzsp.fsf@gitster.g>
In-Reply-To: <220415.868rs6cyal.gmgdl@evledraar.gmail.com>

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

Show 7 quoted lines

> I actually wonder if this wouldn't be that hard, because if we have any
> reflogs we could simply "replay" them by turning (again, I'm a bit rusty
> on the exact lock dance);
>
>     lock(refs/remotes/origin/master);
>     lock(refs/remotes/def/master);
>     create(refs/remotes/def/master, refs/remotes/origin/master^{});

Why deref? Trying to be prepared for seeing a symbolic ref, or something?

Show 13 quoted lines

>     delete(refs/remotes/origin/master);
>     unlock(refs/remotes/origin/master);
>     unlock(refs/remotes/def/master);
>
> Into instead doing:
>
>     lock(refs/remotes/origin/master);
>     lock(refs/remotes/def/master);
>     for from, to, msg ref_update(refs/remotes/origin/master):
>         update(refs/remotes/def/master, from, to, msg);
>     delete(refs/remotes/origin/master);
>     unlock(refs/remotes/origin/master);
>     unlock(refs/remotes/def/master);

Would this (or the above, for that matter) work well when renaming 'foo' to 'foo/bar' (or the other way around), I wonder?

Anyway.

Is the reasoning behind "replay each and every reflog entry" because you are trying to avoid depending on the implementation detail (i.e. "R=.git/logs/refs/remotes; mv $R/origin/master $R/def/master" works only with the files backend)? Unless we are willing to add a new ref backend method to help this natively, it would be a workable but an ultra-slow way to do so, as it would involve open, write, fsync, and close for each reflog entry.

    ... goes and looks ...

Hmph, I am utterly confused. refs/files-backend.c eventually dispatches to files_copy_or_rename_ref() when rename_ref method in ref_storage_be is used on the files backend, and the function already renames the reflog file without having to copy entry by entry. We cannot just open a transaction, run many rename_refs and close it?