Federated training does not happen continuously — it happens in communication rounds. One round is a complete cycle: the server selects clients, sends them the current model, clients train locally, clients send back updates, and the server aggregates. Then the next round begins with the improved model.
Rounds exist because communication is expensive. Uploading from a phone over a metered connection costs battery, bandwidth, and time. If clients reported back after every single training step, the system would spend its life waiting on the network. Batching local work into rounds trades a little training precision for enormous communication savings.
Only a fraction of clients joins each round. With millions of devices available, the server might select a few hundred per round — enough to learn from, few enough to manage. Over many rounds, different subsets participate, so the global model still sees the diversity of the whole population.
A typical training run lasts hundreds or thousands of rounds. Rounds give us a natural unit for measuring progress, cost, and fairness — and they are the loop inside which FedAvg, on Day 6, lives.