Well, I oversimplified a bit. The "tarball" (WARC or ARC file) is a single large file with individually compressed HTTP response objects concatenated together. The index stores the byte offset and length (compressed) into this file, so an HTTP/1.1 range request can pull only the data needed. So in theory it's efficient. I don't work on that team directly and haven't looked in to exactly what the performance latency bottleneck is, sorry for the misleading response.