When to Forget? Complexity Trade-offs in Machine Unlearning