Disable RRLB health-checking for DiscoveryStrategy.ON_NEW_CONNECTION

[idelpivnitskiy]

Motivation:

When DiscoveryStrategy.ON_NEW_CONNECTION mode is used, LB will have a single unresolved entry, the actual IP address will be resolved by a ConnectionFactory. We should not run health-checking logic in this case because it can lead to "all unhealthy" state. We should run a new connection attempt all the time and fail with an exception from transport.

Modifications:

• Configure a custom loadBalancerFactory on a client builder when DiscoveryStrategy.ON_NEW_CONNECTION is used;

Result:

No health-checking for an unresolved address when
DiscoveryStrategy.ON_NEW_CONNECTION is used.

[bryce-anderson]

Code change looks good. Is this worth adding a test?

It looks like the load balancer id is not unique which is the same problem for all load balancers generated by fromDefaults(). Is that something we should address later?

[idelpivnitskiy]

It looks like the load balancer id is not unique which is the same problem for all load balancers generated by fromDefaults(). Is that something we should address later?

The original intend was that id represents who is creating a load balancer factory. When fromDefaults() is used, it means that ST creates it with default settings. When users create their custom LB, they suppose to provide a different id so that providers will see the difference between the "default one" and a "custom one".

I'm agreed that there could be use-cases, when users need to change an LB config for a specific client, even if it was fromDefaults() originally. In this case, we need something more than just id, maybe a bundle of id and targetResource. We can keep this in mind for the new LoadBalancerBuilder you are working on. Not sure if there is currently a use-case for RoundRobinLoadBalancer.

Is this worth adding a test?

Good call, will add one.