Aurora client should not automatically retry non-idempotent operations

Aurora client has a built in mechanism to automatically retry thrift API operations if the connection with scheduler times out, experiences transport exception, or encounters a transient exception on the scheduler side.

Retrying thrift calls due to scheduler connection timeout and transient exceptions (see AURORA-187) is safe. However, as Aurora has no concept of idempotency, its client can retry non-idempotent operations upon encountering transport exceptions which can lead to nondeterministic situations.

For example, if client requests go through a proxy to reach scheduler, client might consider a non-idempotent request failed and automatically retry it while the original request has been received and processed by the scheduler.