I suspect there are other places that could be changed from doing a one-shot compression to using a streaming compressor, but I admit that I didn't look too carefully. This was really just to get over the hump of running out of memory so that I could continue to debug other issues.

I do have interest in looking closer at other zlib usages, but unfortunately I don't have the time to do that at present.

This looks really nice to me. The zstream stuff seems like a better abstraction level and I agree that we can look later at other places that can be improved similarly. I have a couple of questions:

1. Do you think it is worth trying to avoid the repeated zbuf alloc/free inside write_object and move that temporary buffer into the packbuilder? Perhaps relative to everything else that has to be done, the allocation is too trivial to matter, but since it is a fixed size buffer, it seems like it could be allocated once and then just reused until we are done.
2. Along those lines, is there any point in adding a git_zstream_reset function (that calls zlib's deflateReset function) so that we don't have to repeated recreate and free the zstream object? I don't know what the overhead is, but again it could be factored out if it made a performance difference.

Obviously both of those questions can be answered at a future point if you just want to get this in as an incremental improvement. The code looks good to me! 🤘

Yes, that definitely makes sense. Let's sit on this then until I have a few minutes to look at that.

Rebased and updated with @arrbee 's fine suggestions.

Rrrrrrebased! Any objections? If not, I shall merge this bad boy.