In this paper, we use a New Radio (NR) simulator, based on ns-3, to assess the impact of 5G NR numerologies on the end-to-end (E2E) latencies in a realistic and complex scenario, including TCP and UDP flows. As expected, we found that TCP goodput increases with the numerology, since a larger numerology allows reducing the round-trip-time. However, although counter-intuitive, simulation results exhibit that the E2E latency of uplink (UL) UDP flows may not be reduced with the numerology. In fact, it depends on two key factors and their relationship: the processing delays (fixed or numerologydependent) and the inter-packet arrival time, which depends on the UDP flow rate and the packet size. We demonstrate how, in some cases, the latency is worsened by an increasing signaling exchange that grows with the numerology. In particular, this is due to a handshake mechanism in UL (scheduling request and UL grant) that is performed each time a data packet encounters empty RLC buffers. For some combination of flow rate, packet size, and processing delays that are not numerology dependent, increasing the numerology may not reduce the E2E delay. Therefore, we conclude that the selection of the numerology in an NR system should be carefully made by taking into account the traffic patterns and the processing delays.