Federated Learning (FL) incurs high communication overhead, which can be\ngreatly alleviated by compression for model updates. Yet the tradeoff between\ncompression and model accuracy in the networked environment remains unclear\nand, for simplicity, most implementations adopt a fixed compression rate only.\nIn this paper, we for the first time systematically examine this tradeoff,\nidentifying the influence of the compression error on the final model accuracy\nwith respect to the learning rate. Specifically, we factor the compression\nerror of each global iteration into the convergence rate analysis under both\nstrongly convex and non-convex loss functions. We then present an adaptation\nframework to maximize the final model accuracy by strategically adjusting the\ncompression rate in each iteration. We have discussed the key implementation\nissues of our framework in practical networks with representative compression\nalgorithms. Experiments over the popular MNIST and CIFAR-10 datasets confirm\nthat our solution effectively reduces network traffic yet maintains high model\naccuracy in FL.\n