Client Availability in Federated Learning: It Matters!

Achieving efficient Federated Learning (FL) in large-scale cross-device deployments is challenging, in part due to stragglers and stale updates. However, another key yet often overlooked challenge is client unavailability. Some clients selected for training remain inactive and fail to participate. Unlike stragglers, which eventually return updates, unavailable clients provide no updates, causing rounds to stall. Even when they do return, they introduce significant delays, both of which increase time-to-accuracy. Existing FL approaches, including synchronous FL (SyncFL) and asynchronous FL (AsyncFL), fail to address client unavailability- SyncFL suffers from long waits, while AsyncFL experiences increased update staleness. Some recent client selection strategies assume oracular knowledge of client availability and prioritize historically high-utility clients. Not only is this an unrealistic assumption, but these strategies also rely on stale client utility values, which become increasingly stale as client unavailability rises. To enable scalable and efficient FL in real-world scenarios, we argue that it is crucial to address fluctuating and high client unavailability. This work highlights its impact on FL and underscores the need for a real-time, availability-aware mechanism that improves client selection and ensures high-quality update aggregation, ultimately reducing time-to-accuracy.