The OnHandshake RPCs start too early. They are started before the TCP connection is even established (when the stub is created, which is part of newHandshaker). If the connection is never established (e.g., the server can't be reached), the TsiHandshakeHandler won't be added to the Channel and so the RPC cleanup will never fire. Before the timeout added in #7589, this would be very bad.

We need to fix the bug, and evaluate how exposed our existing clients are. The v1.34 release is semi-okay, since it has a timeout in place so these busted RPCs won't live forever. But in the case that all backend connections are failing, it might still produce ugly results.

CC @veblush, @apolcyn