On Thu, Sep 30, 2021 at 11:31:32PM -0700, Elijah Newren wrote: > * Do something different with the tmp-objdir usage or use > pretend_object_file(). I'm slightly worried the latter was only > designed for a few files that all fit into memory, which may not be > valid for this usecase. Just to make sure I was being clear on this final point: the pretend_object_file() interface _now_ is definitely not suitable for the sizes you'd want. I think you'd need to: - have it handle pointers to non-duplicated storage, instead of owning all buffers (and then the remerge-diff code would do something clever with mmaps and tempfiles). - use something better than the linear cached_objects array for storing the entries, so lookup doesn't suffer with many objects. Perhaps an oidmap. - teach it to remove objects. For your purposes, a complete clear function would be sufficient. -Peff