A common pattern in libraries doing I/O is to receive data in chunks, put them in a list, then join them all together using b"".join(chunks). For example, see http.client.HTTPResponse._safe_read. When the output is large, the memory copies can block the interpreter for a non-trivial amount of time, and prevent multi-threaded scaling. If the GIL could be dropped during the memcpys it could improve parallel I/O performance in some high-bandwidth scenarios (36050 mentions a case where I've run into this serialisation bottleneck in practice).

Obviously it could hurt performance to drop the GIL for small cases. As far as I know numpy uses thresholds to decide when it's worth dropping the GIL and it seems to work fairly well.