Retry is already implemented (except hedging support) in v1.16.1, but there is no other document than the spec yet, because there are some caveats for the time being: (1) users need manage to disable census stats and tracing when enabling retry,(2) there are some caveats for using service config. In next release, v1.17.0, enabling retry will automatically disable census.
As for DEADLINE_EXCEEDED, it does not make sense to retry at the library level, because a retry attempt would immediately exceed the deadline of the RPC. Users could implement application level retry for DEADLINE_EXCEEDED by themselves by making a new RPC.