Towards Understanding the Feasibility of Machine Unlearning