The master process opens
--http-port/-H, allowing workers to accept connections on it, as they are forks.
If all workers are busy, this connection is put into a system backlog and will be handled once any worker becomes free. This backlog has a system-defined maximum size, and when overflowed, all new connections would be denied. For an external observer (nginx in our case), sending a request to all-workers-busy-KPHP seems like “takes a long time to connect”, so you can configure connection timeout to handle this situation somehow.
For instance, if you have many KPHP-backends and one or many nginx-frontends (on separate servers), you can use
proxy_next_upstream for balancing between KPHP.