git push a huge repository to a server with limited memory - out-of-memory

Question

The server has only 64MB of memory. I'm trying to push a huge git repository to it. Initially the target directory contains an empty bare repository. The push fails:
$ git push server:/tmp/repo master
Counting objects: 3064514, done.
Compressing objects: 100% (470245/470245), done.
fatal: Out of memory, calloc failed
error: pack-objects died of signal 13
error: failed to push some refs to 'server:/tmp/repo'
$ ssh server cat /tmp/repo.git/config
[pack]
threads = 1
deltaCacheSize = 8m
windowMemory = 32m
[core]
repositoryformatversion = 0
filemode = true
bare = true
I get the same error message after changing git config pack.windowMemory 16m on the server.
The same push succeeds to localhost:
$ git push 127.0.0.1:/tmp/repo master
Password:
Counting objects: 3064514, done.
Compressing objects: 100% (470245/470245), done.
Writing objects: 100% (3064514/3064514), 703.02 MiB | 10.84 MiB/s, done.
Total 3064514 (delta 2569775), reused 3059081 (delta 2565342)
To 127.0.0.1:/tmp/repo
* [new branch] master -> master
Is there a remote git config setting which can make the push succeed? Or do I have to repack the repo locally before pushing (with what settings)?
Please note that using a different server with more memory is not an option. Adding memory to the existing server is an option, up to 96MB. It's OK for me to use more disk space than usual on the server if the memory limit is met.
Similar question without a working solution: https://serverfault.com/questions/372899/git-fails-to-push-with-error-out-of-memory
Repacking the repository locally didn't help, git push prints the same error. Repack settings in the local repo:
git config core.packedgitlimit 32m
git config core.packedgitwindowsize=32m
git config pack.threads 1
git config pack.deltacachesize 8m
git config pack.windowmemory 32m
git config pack.packsizelimit 500m
My idea is that the reason why it fails is that the total number of objects is too large: even the SHA-1 hashes won't fit (20 * 3064514 bytes is almost 64MB).

Answer 1

Possible other causes
As #torek pointed out in his comment, this may not be an indication of the server running out of memory, but an indication that something is going wrong locally. Perhaps something changed between when you were pushing to server and to local host that freed up memory on your local machine?
It's also plausible that git is figuring out that you're pushing to localhost, and bypassing the "Git aware" transport mechanism and/or using hardlinks, which might reduce the memory needed. I don't see any indication in the docs that it WOULD do this, and I'm not sure off the top of my head how you could test this, or force it not to do that, but it's a possibility.
Another possible issue is that the host.xz:path/to/repo.git/ url syntax is only recognized if there are no slashes before the first colon, so depending on what server is, that could be causing problems.
If none of these are the case, and the problem is in fact that it's running out of memory on the server, you might have a few options here, depending on the circumstances. I don't know if any of these will work, but they're worth a try.
Solution 1: don't push all the commits at once
I'm assuming you've got many commits in the commit history of master. Try pushing them in stages. E.g.
git push server:/tmp/repo master~500
git push server:/tmp/repo master~400
git push server:/tmp/repo master~300
git push server:/tmp/repo master~200
git push server:/tmp/repo master~100
git push server:/tmp/repo master
Solution 2: Push individual objects one at a time
This is going to be incredibly tedious and DEFINITELY need to be automated/scripted on your local machine. However, you don't actually need to push whole commits all at once.
Instead, you can push individual objects one at a time as long as you push them to a tag ref instead of a branch ref. E.g. if we were working with https://github.com/llvm/llvm-project and wanted to push the tree object 0082ee0b3ad78ff55b2a3a65ef5bfdb8cd9713a1 from it (this is the tree object pointed to by commit faf5e0ec737a676088649d7c13cb50f3f91a703a), we could do git push server:/tmp/repo 0082ee0b3ad78ff55b2a3a65ef5bfdb8cd9713a1:refs/tags/test. Using this we can push individual objects one at a time, starting with blobs, then the tree objects, then finally commit objects. We'd end up with a TON of tags to clean up later, but I'll leave that to you to figure out.
For the rest of these solutions, I'm working under a couple assumptions:
Given the limitations you described, and the way you specified the url as server:/tmp/repo instead of something ending with .git, I'm assuming this remote repository isn't going to be managed with any service like github or gitlab, which should give you a little more room to use some unconventional techniques.
I'm also assuming you probably have the ability to log on to/run
commands on the server.
If either of these are not the case, and the above didn't work, I'm out of ideas at the moment.
Solution 3: backwards push using fetch or clone
There's actually nothing special about a server, it's just another git repository that you can trade commits with. The only difference is that a server is usually hosting what's called a bare repository: it doesn't typically keep a working tree of it's own (in other words, it only keeps the contents of the .git folder).
So, try performing the push in reverse using fetch/clone from server:
Push to a third, intermediate server (let's call it server2). Ideally, one with a lot more performance, like a github hosted repo.
Log onto/ssh into server, and from there, clone the repo into /tmp/repo: git clone --bare git#github.com:path/to/your/repo.git.
I would be surprised if this solved anything on it's own, but it's worth trying and step 1 will still set us up for solution 4 and 5. If by chance it does work, you can tidy up by removing server2 as a remote on server: git remote remove origin, then setting up your remotes on your local machine to point towards server instead of server2.
Solution 4: backwards push, but without fetching all the commits at once
Like solution 3, push to an intermediate server, but this time, instead of using clone and fetching everything all at once, fetch the commits in stages:
Log onto/ssh into server, and from there, initialize /tmp/repo as a bare repo:
cd /tmp/repo
git init --bare
git remote add origin git#github.com:path/to/your/repo.git
Still on server, fetch commits one at a time:
git fetch origin 569d84fe99e63e830ea036598f7fa7a5f9899d7c
git fetch origin 9aaba9d9bb4fc3648a9417820858086b14b6b73e
git fetch origin faf5e0ec737a676088649d7c13cb50f3f91a703a
Solution 5: backwards push, but using partial and/or shallow clones
Instead of fetching individual commits, we can use partial and or shallow clones to restrict how much we are fetching at once. There is a good write-up explaining what those are on the github blog: https://github.blog/2020-12-21-get-up-to-speed-with-partial-clone-and-shallow-clone/. This time, we can't won't use a bare repository. We want to be able to check out commits to fill in the missing objects later. You can follow the instructions here to convert it to a bare repository when you're done. Alternatively, instead of using a regular (non-bare) repository, explicitly fetching the objects might also work, but I don't know for sure off the top of my head.
I think everything I've already written, combined with that write-up should give you all the pieces you need to figure out how to do this. I've already spent hours writing this up, it's late, this solution's kind of complicated, and it's an esoteric question that hasn't been touched in years. If somebody comes across this and needs a more complete answer for this, leave a comment and I'll fill it in, but this is as far as I'm willing to go right now for some potential internet points if nobody actually needs this answer XD.