Because this code was doing pointer equality checks, it would work for
the case of a failed attempted RPC because the objects are from the
manager itself:
https://github.com/hashicorp/consul/blob/v1.0.3/agent/consul/rpc.go#L283-L302
But the pointer check would always fail for events coming in from the
Serf path because the server object is newly-created:
https://github.com/hashicorp/consul/blob/v1.0.3/agent/router/serf_adapter.go#L14-L40
This means that we didn't proactively shift RPC traffic away from a
failed server, we'd have to wait for an RPC to fail, which exposes
the error to the calling client.
By switching over to a name check vs. a pointer check we get the correct
behavior. We added a DEBUG log as well to help observe this behavior during
integrated testing.
Related to #3863 since the fix here needed the same logic duplicated, owing
to the complicated atomic stuff.
/cc @dadgar for a heads up in case this also affects Nomad.
Josh Soref
James Phillips
Kyle Havlovitz
Frank Schroeder