"git fetch --refetch" and multiple (separate/orphan) branches

Hi folks,

I just recently noticed that "--refetch" was added in 2.36, and I got pretty excited - the ability to "fill in" missing blobs after a too-filtered clone is something that I've wanted a number of times, as I mentioned in 2021 in thread https://public-inbox.org/git/CAPMMpohOuXX-0YOjV46jFZFvx7mQdj0p7s8SDR4SQxj5hEhCgg@mail.gmail.com/ .

When I first ran "git fetch --refetch" today however (git 2.38.1, against server git/2.38.4.gl1), with a configured blob filter of "blob:1100M", a much higher size than any blob in the history, it only got a *relatively* small number of objects - 3GB of data rather than the 18GB that a new unfiltered fetch would have retrieved.

After some more testing I tried again, and got the expected outcome that time. The relevant difference between the two attempts is that in the first case, when I only got some of the objects I expected, there was an updated tag as a result of the fetch. The second time, when I got everything, there were no updated refs.

In this repository there are several "independent" sets of branches, and the tag updated in that first fetch belongs to one of the smaller-history branches.

What I believe is happening is that *if* there are refs to be updated (or new refs, presumably), *then* the objects returned to the client are only those required for those refs. If, on the other hand, there are no updated refs, then you get what is advertised in the doc: "all objects as a fresh clone would [...]".

I've tested a couple of different scenarios and the behavior seems consistent with this explanation.

In a repo where all branches are derived from the same history, this probably isn't very noticeable; in the repo I'm working on it makes a huge difference, so the only way I can imagine getting "correct" behavior would be to always to a "git fetch" right before the "git fetch --refetch".

Is this a bug, or expected behavior that should be noted in the doc, or do we consider the multiple-independent-branches usecase to be edge-casey enough to be an easter egg for people like me?

Thanks, Tao

Hi Tao

On Fri, 2 Jun 2023 at 22:23, Tao Klerks <tao@klerks.biz> wrote:

Do you have a repo & steps that could reproduce this easily? Otherwise I can try and work up something.

At first glance it appears to be a bug.

Thanks,

Rob :)

Hi Robert,

Sorry about the extended delay, I haven't had a chance to do "git hacking" in a while.

On Sat, Jun 3, 2023 at 10:18 AM Robert Coup <robert@coup.net.nz> wrote:

Does the following work? It shows that with a change to the orphan branch from another client, a refetch in the original client gets about half the objects (the ones for the orphan branch that was updated), and in another fetch right after, with no new changes, the refetch gets all 600-or-so objects.

create_n_commits() {
  for i in $(seq $2); do
    echo "another new line $RANDOM" >> "$1/datafile"
    git -C "$1" add datafile
    git -C "$1" commit -m anothercommit -q
  done
}

mkdir refetch-testing SERVERFOLDER=refetch-testing/server git init "$SERVERFOLDER" --bare

CLIENTFOLDER=refetch-testing/client git init "$CLIENTFOLDER" git -C "$CLIENTFOLDER" remote add origin "../server"

git -C "$CLIENTFOLDER" checkout -b main create_n_commits "$CLIENTFOLDER" 100 git -C "$CLIENTFOLDER" push origin HEAD

git -C "$CLIENTFOLDER" checkout --orphan orphan create_n_commits "$CLIENTFOLDER" 100 git -C "$CLIENTFOLDER" push origin HEAD

echo "---HERE IS A NORMAL FULL REFETCH---" git -C "$CLIENTFOLDER" fetch --refetch echo "---NORMAL FULL REFETCH ENDS---"

OTHERCLIENTFOLDER=refetch-testing/otherclient git clone "$SERVERFOLDER" "$OTHERCLIENTFOLDER" git -C "$OTHERCLIENTFOLDER" checkout orphan create_n_commits "$OTHERCLIENTFOLDER" 5 git -C "$OTHERCLIENTFOLDER" push origin HEAD

echo "---HERE IS A WEIRD PARTIAL REFETCH OF ONE BRANCH ONLY---" git -C "$CLIENTFOLDER" fetch --refetch echo "---WEIRD PARTIAL REFETCH ENDS---"

echo "---HERE IS NORMAL REFETCH AGAIN---" git -C "$CLIENTFOLDER" fetch --refetch echo "---NORMAL REFETCH ENDS---"

rm -rf refetch-testing