One overlooked point is ephemeral source port exhaustion. If a load balancer forwards a HTTP connection to a backend system, it needs a TCP source port for the duration of that connection (not destination port, which is probably 80 or 443). That limits the number of outgoing connections to less than 65535. A common workaround is to use more outgoing IP addresses to the backends as source IPs, thus multiplying the available number of source ports to 65535 times number_of_ips.
HTTP/2 solves this, as you can multiplex requests to backend servers over a single TCP socket. So there is actually a point of using HTTP/2 for load_balancer <-> backend_system connections.
HTTP/2 solves this, as you can multiplex requests to backend servers over a single TCP socket. So there is actually a point of using HTTP/2 for load_balancer <-> backend_system connections.