Right now we use a generation index to processFailure() for connection pooling. This has two unintended side effects in spurious networking conditions:

• A failed connection causes connections created after it to be considered failed as well. This could throw away good connections in situations with spurious failures.
• A failed connection causes a connection request (and thus an operation) to receive a not okay status. This can happen even if the remote host has already failed the connection by the time the request came in.

We should use either a monotonic sequence id or clock to mark each time a connection is considered "ready" (including initial creation). We should also attach these ids to requests. When a connection fails, all ids before it should be considered failed.