We study analytically and experimentally the throughput of the packetized time-varying discrete erasure channel with feedback, which closely captures the behavior of many practical physical layers. We observe that the channel variability at different time scales affects the link-level throughput positively or negatively depending on its time scale. We show that the increased variability in the channel at a time scale smaller than a single packet increases the link-level throughput, whereas the variability at a time scale longer than a single packet reduces it. We express the throughput as a function of the number of transmissions per packet and evaluate it as in terms of the cumulants of the samples of the stochastic processes, which model the channel. We also illustrate our results experimentally using mote radios.