This is my guess - two write calls to one write call. I think this will happen to work even in the presence of Nagle’s algorithm, because there is initially not going to be an already sent packet waiting acknowledgement, so at least the first write call will probably be unbuffered in that sense.
Assuming an mtu of 1500, ~40 byte TCP header, and a little bit of HTTP/1.1 headers, that leaves probably plenty of room for the etcd PUT to not get fragmented, I think.