The control-plane components rely on client-go to connect to the Kubernetes API. Issues like #2165 and #2182 highlight the need for better visibility into that client, to confirm the client is healthy, with a consistent view of Kubernetes' state.

We should look for ways to instrument the Kubernetes client with at least the following:

• request/response counts
• request errors
• timeouts
• latency

Prometheus has middleware wrappers for clients, that may be a good starting point.