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

["Ævar Arnfjörð Bjarmason" <avarab@gmail.com>]

On Wed, Apr 13 2022, brian m. carlson wrote:

> [[PGP Signed Part:Undecided]]
> In my day-to-day work, I have the occasion to use GitHub Codespaces on a
> repository with about 20,000 refs on the server.  The environment is set
> up to pre-clone the repository, but I use a different default remote
> name than "origin" ("def", to be particular), and thus, one of the things
> I do when I set up that environment is to run "git remote rename origin
> def".

Aside from how we'd do renames with transactions, do you know about
clone.defaultRemoteName and --origin?

> This process takes 35 minutes, which is extremely pathological.  I
> believe what's happening is that all of the refs are packed, and
> renaming the ref causes a loose ref to be created and the old ref to be
> deleted (necessitating a rewrite of the packed-refs file).  This is
> essentially O(N^2) in the order of refs.
>
> We recently added a --progress option, but I think this performance is
> bad enough that that's not going to suffice here, and we should try to
> do better.
>
> I found that using "git for-each-ref" and "git update-ref --stdin" in a
> pipeline to create and delete the refs as a single transaction takes a
> little over 2 seconds.  This is greater than a 99.9% improvement and is
> much more along the line of what I'd expect.
>
> I thought about porting this code to use a ref transaction, but I
> realized that we don't rename reflogs in that situation, which might be
> a problem for some people.  In my case, since it's a freshly cloned repo
> and the reflogs aren't interesting, I don't care.

There was a (small) thread as a follow-up to that "rename --progress"
patch at the time, did you spot/read that?:
https://lore.kernel.org/git/220302.865yow6u8a.gmgdl@evledraar.gmail.com/

There's doubtless other previous discussions, I just haven't
found/remember them.

I have (briefly) tried hacking on this myself in the past, as anyone
who'll poke at that will no doubt find "branch rename" and "branch copy"
non-ref-transaction way of doing this are basically other callers with
the same problem.

Before I go any further I think it's good to know how far down this
particular rabbit hole you already are...

> I think a possible way forward may be to either teach ref transactions
> about ref renames, or simply to add a --no-reflogs option, which omits
> the reflogs in case the user doesn't care.  I'm interested to hear ideas
> from others, though, about the best way forward.

More generally, probably:

 1. Teach transactions about N operations on the same refname, which
    they'll currently die on, renames are one case.

 2. Be able to "hook in" to them, updating reflogs is one special-case,
    but we have the same inherent issue with updating config in lockstep
    with transactions.