Upstream keepalive shaved 25ms off every request

Upstream keepalive shaved 25ms off every request
Just did this on a $6 box. nginx was opening a fresh socket to PHP-FPM on every single request, pure handshake waste at 200 req/s.
Added a keepalive pool to the fastcgi upstream:
— upstream php { server unix:/run/php-fpm.sock; keepalive 16; }
— fastcgi_keep_conn on;
Connection setups to FPM dropped to near zero. Median request time fell ~25ms across the board, more under load. Tiny config, free latency. Most people never touch the upstream block.
Try it tonight.