Federated Learning (FL) has become a popular paradigm for learning from distributed data. To effectively utilize data at different devices without moving them to the cloud, algorithms such as the Federated Averaging (FedAvg) have adopted a "computation then aggregation" (CTA) model, in which multiple local updates are performed using local data, before sending the local models to the cloud for aggregation. However, these schemes typically require strong assumptions, such as the local data are identically independent distributed (i.i.d), or the size of the local gradients are bounded. In this paper, we first explicitly characterize the behavior of the FedAvg algorithm, and show that without strong and unrealistic assumptions on the problem structure, the algorithm can behave erratically for non-convex problems (e.g., diverge to infinity). Aiming at designing FL algorithms that are provably fast and require as few assumptions as possible, we propose a new algorithm design strategy from the primal-dual optimization perspective. Our strategy yields a family of algorithms that take the same CTA model as existing algorithms, but they can deal with the non-convex objective, achieve the best possible optimization and communication complexity while being able to deal with both the full batch and mini-batch local computation models. Most importantly, the proposed algorithms are {\it communication efficient}, in the sense that the communication pattern can be adaptive to the level of heterogeneity among the local data. To the best of our knowledge, this is the first algorithmic framework for FL that achieves all the above properties.