See siderolabs/talos#7191

Current implementation: round-robin across all live backends.

Desired implementation (optional, enabled as an option):

• backend healtcheck returns a "rank" for all healthy backends
• picking up a backend should start with the lowest "rank" - round-robin across live backends with the same rank
• if the "rank" is empty, proceed to the next rank

What rank can be: e.g. healthcheck does TCP connect, measures latency, and splits in log_10(t) buckets: [<1ms, <10ms, <100ms, <1000ms, ...], which are mapped to ranks 0, 1, 2, 3, ...

The end goal is to make sure that loadbalancer picks the backends which are "better"/"closer"/"lower latency".

See

Add a tier which will cover <0.1ms (for local node addresses).