Standard Federated Learning (FL) operates under a comforting fiction: a stable, predictable pool of clients. In textbook implementations of Federated Averaging (FedAvg), a central server samples uniformly from a fixed set of devices, aggregates their local updates, and repeats the process until convergence.
But real-world edge networks—composed of mobile phones, autonomous vehicles, and IoT sensors—are fundamentally chaotic. Devices drop offline when disconnected from Wi-Fi, enter sleep modes to save battery, or move across geographic boundaries. As outlined in A Survey on Federated Learning Systems: Vision, Hype and Reality for Data Privacy and Protection by Qinbin Li et al., this high degree of participant autonomy and constant churn distinguishes practical federated systems from traditional managed distributed computing. When the client pool is a revolving door, standard convergence guarantees break down.
Why Client Churn Destabilizes Learning
To understand why dynamic participation breaks standard FL, we must look at how it distorts the optimization landscape.
First, client availability is rarely independent and identically distributed (i.i.d.). Instead, it is highly correlated with real-world factors like time zones, local charging habits, or network conditions. As Youbang Sun et al. point out in On the Stability Analysis of Open Federated Learning Systems, ignoring these regional correlations can severely destabilize the learning process. If only clients from a specific time zone are awake during certain training rounds, the global model overfits to their local data distributions, causing the global objective itself to shift over time.
Zhenyu Sun et al. formalize this in Debiasing Federated Learning with Correlated Client Participation by modeling non-uniform participation as a Markov chain. When clients must wait a minimum number of rounds before rejoining training due to system constraints, standard FedAvg accumulates an unavoidable statistical bias. The global model no longer tracks the true average of the entire population; it drifts toward the objectives of the most readily available or frequently active devices.
Tracking Drift and Debiasing Updates Adaptively
How do we correct for this bias when the server does not know the total number of participants or who will show up next?
The intuition behind modern solutions is to make the optimization process agnostic to the absolute size or stationarity of the client pool. For instance, in AdaBest: Minimizing Client Drift in Federated Learning via Adaptive Bias Estimation, Farshid Varno et al. introduce a method to accumulate local observations into a global drift estimate without requiring prior knowledge of the total participant population. By maintaining a running stabilization factor, the server can smooth out the erratic updates caused by unpredictable client availability.
Alternatively, the burden of debiasing can be shifted to the clients themselves. The algorithm proposed by Zhenyu Sun et al. allows clients to dynamically estimate their own long-term participation rates based on local historical availability. When selected, a client scales its model update inversely proportional to its participation probability. This simple re-weighting ensures that rare or intermittent clients contribute with a higher magnitude when they do appear, mathematically steering the global model back toward the true global optimum.
Warm-Starting and Rapid Adaptation
When entirely new clients join an open system mid-training, they often bring novel data distributions. This sudden influx can shock the global model, causing it to forget previously aggregated knowledge—a problem exacerbated in class-incremental settings.
To mitigate this, Dong-Jun Han et al. propose a "plug-and-play" initialization framework in Federated Learning with Dynamic Client Arrival and Departure: Convergence and Rapid Adaptation via Initial Model Construction. Instead of forcing a newly arrived population to immediately synchronize with a rigid global model, the server uses gradient similarity to construct a tailored initial model for the incoming cohort. By warm-starting the training process based on how closely the new clients' data aligns with historical updates, the system achieves massive convergence speedups and prevents the global model from destabilizing during population shifts.
Just-in-Time Scheduling and Online Recruitment
In a static system, participant selection happens upfront. In a dynamic system, the server must make immediate, "online" decisions as devices flicker in and out of availability.
Ihab Mohammed highlights in Budgeted Online Selection of Candidate IoT Clients to Participate in Federated Learning that traditional stateful selection mechanisms fail in stateless environments where participants must be re-selected every round. To address this, auction-based and market-driven frameworks have shifted from offline bidding to online, multi-round recruitment.
For example, Xavier Tan et al. leverage Lyapunov optimization in Hire When You Need to: Gradual Participant Recruitment for Auction-based Federated Learning to gradually recruit data owners over multiple rounds. This "just-in-time" recruitment allows the server to dynamically adjust the composition of the training pool based on evolving model needs and real-time budget constraints, rather than over-committing to a fixed set of contributors at the outset.
[Server] ──(Predicts Availability via Semi-Markov Model)──┐
│ ▼
├──► [Client A: High Reliability] ──► Scheduled for Immediate Training
└──► [Client B: Flaky/Low Battery] ──► Deferred to Avoid Network Stalls
Furthermore, predicting when a client will disappear is crucial for semi-decentralized topologies where edge servers aggregate local updates. In TRAIL: Trust-Aware Client Scheduling for Semi-Decentralized Federated Learning, Gangqiang Hu et al. employ an Adaptive Hidden Semi-Markov Model to forecast how long a volatile client will remain reliable. By proactively scheduling client-server associations based on these temporal availability predictions, the framework prevents "flaky" participants from stalling the network or degrading model convergence.