It looks like for localhost buffers are bigger, but even if it's not local host I do get 5MB stuck in socket buffers. I was only able to get perfect results by writing my own proxy in c and doing some obscure nodejs code to avoid buffering.

In any case, if nginx does not provide a way to control sockets buffer I cannot use it. For example, 'X-Accel-Buffering: no' supposedly disables caching (I didn't see any effect of that anyways), so I wanted to add some kind of headers to be able to tell nginx what buffers to set per connection. In my case I do regular reverse proxy stuff with nginx, but on certain connections I need exact control and nginx doesn't provide any of that. haproxy for example worked much better for me, but it's sndbuf/rcvbuf settings are global, which is equally unacceptable.

Would that be easy to add headers like these X-Accel-Up-Rcvbuf: 12345, X-Accel-Down-Sndbuf: 4567 so that on getting them from upstream nginx would configure sockets that are used by that connection?