For that, we would need to stream the TOC and tar-split data to/from disk, never storing all of it in memory.

Currently, per #2230 , if we hit the limit, we fall back to traditional pulls.