Swap to 'fail fast' and 'ignore connectivity' behavior

[ejona86]

The other implementations are swapping to fail fast by default. We believe this is what users would typically assume our behavior to be and it behaves well with infinite deadlines, as well as being required by many latency-sensitive applications. These are not the same semantics as in Stubby, because gRPC will delay the RPC while in IDLE or CONNECTING until the state becomes READY or TRANSIENT_FAILURE.

Java today is doing neither 'fail fast' nor 'ignore connectivity', but instead something in-between because it delays while in TRANSIENT_FAILURE, but only one connection attempt.

This is a very important behavior change for users and must be done before GA.

[ejona86]

To be more clear, swapping to 'fail fast' by default has to happen before GA. Implementing non-fail fast could happen after GA, but it would be nice for GA.

[ejona86]

For fail fast, this change primarily involves tweaking TransportSet's behavior. For specifying non-fail fast it looks like it may be harder than I originally thought, because there will need to be some special plumbing and it's not obvious to me yet the appropriate design. It would need plumbing via CallOptions to allow the client to specify the behavior, but that isn't enough by itself.

[ejona86]

Channel states are defined in this spec.