The library makes one record per call (up to 8 KiB of data); a write of more goes through several
records into out and reaches the connection in one send, up to this much, instead of one send
(and one trip through the scheduler) per record.
BATCH_BYTES :: 64 * 1024