We consider the joint dynamic assortment and pricing problem under the multinomial logit model (MNL) with an unknown time-varying Poisson arrival rate and preference for customers. In the model, for each period, the seller chooses the assortment and product prices jointly to maximize her revenue. The arrival rate of the customers depends on the assortment and prices offered with unknown parameters. We propose an efficient algorithm where the seller can choose assortment and product prices jointly to learn the underlying parameters about the Poisson arrival and preference based on the realized arrival and choices of customers. We show that our algorithm is asymptotically efficient in the sense that the regret is bound with \sqrt{T}\log(T) with a high probability 1-O(1/T) and provide a matching lower bound (up to \log(T)) showing the optimality of our algorithm.