Adaptive Data Admission and Retention for Streaming Federated Learning

A joint server-side admission and client-side memory-management framework for streaming federated learning under sampling-cost and buffer constraints

Zhuoyi Zhao; Ben Liang· 2026· DOI 10.48550/arXiv.2607.23987

The core problem

Streaming federated learning (FL) faces a fundamental tension: clients continuously generate new training data, but their memory buffers are finite, and sampling each data point incurs a time-varying cost. The paper addresses the problem of selectively admitting and retaining data over time to minimize the cumulative excess population risk under a sampling-cost budget and buffer constraints. The authors consider a joint server-side admission and client-side memory-management framework. The core challenge is to balance the learning benefit of fresh data against the cost of sampling and the risk of buffer overflow, all while operating in a decentralized setting where clients cannot share raw data. The work derives a learning-error bound that explicitly captures the effects of instantaneous training sample size, distinct-sample growth, and reuse imbalance through a characterization of the effective sample size. This bound motivates a surrogate penalty that guides the design of an online policy.

The authors provide theoretical guarantees for ACDPP. Through a sequence of comparison arguments via an auxiliary constant-admission policy, they connect the ACDPP learning bound to a costless oracle benchmark. This yields explicit guarantees in terms of sublinear regret and sampling-cost violation. Specifically, the regret is O (T), and the sampling-cost violation is also sublinear, controlled by the time-varying rectangular admission region. The buffer-occupancy violation is controlled through offline selection of the retention horizon K . The theoretical analysis shows that ACDPP achieves a trade-off between learning performance and constraint satisfaction. Experiments on multiple datasets demonstrate that the proposed policy remains close to the oracle benchmark while satisfying the sampling-cost and buffer constraints. The empirical results confirm the theoretical findings, showing that ACDPP outperforms baseline policies that do not adapt to time-varying costs or buffer constraints.

Streaming federated learning (FL) faces a fundamental tension: clients continuously generate new training data, but their memory buffers are finite, and sampling each data point incurs a time-varying cost. The paper addresses the problem of selectively admitting and retaining data over time to minimize the cumulative excess population risk under a sampling-cost budget and buffer constraints. The authors consider a joint server-side admission and client-side memory-management framework. The core challenge is to balance the learning benefit of fresh data against the cost of sampling and the risk of buffer overflow, all while operating in a decentralized setting where clients cannot share raw data. The work derives a learning-error bound that explicitly captures the effects of instantaneous training sample size, distinct-sample growth, and reuse imbalance through a characterization of the effective sample size. This bound motivates a surrogate penalty that guides the design of an online policy.

The authors formulate the problem as minimizing the cumulative excess population risk over a time horizon T, subject to a sampling-cost budget and per-client buffer constraints. Let x_{t} denote the admission decision at time t (1 if admitted, 0 otherwise), and let c_{t} be the time-varying sampling cost. The buffer occupancy evolves as B_{t + 1} = max (0, B_{t} + x_{t} - r_{t}), where r_{t} indicates whether a sample is retained. The learning-error bound is expressed in terms of the effective sample size N_{eff}, which depends on the instantaneous training sample size, the growth of distinct samples, and the reuse imbalance. Specifically, the excess risk is bounded by O (1/ N_{eff}) . To minimize this bound under the constraints, the authors introduce a surrogate penalty and develop the Active-Constraint Drift-Plus-Penalty (ACDPP) policy. ACDPP combines a structured client-side K -step retention rule with a server-side online admission rule and a time-varying rectangular admission region. The retention rule keeps samples for at most K steps, while the admission rule decides whether to sample based on the current cost and the buffer state. The rectangular admission region is defined by time-varying thresholds that ensure the sampling-cost budget is respected. The policy is online and does not require future knowledge of costs or data arrivals.

The paper's key contribution is the derivation of a learning-error bound that explicitly accounts for the effects of instantaneous training sample size, distinct-sample growth, and reuse imbalance. This bound provides a principled way to design online policies for streaming FL with limited memory. The ACDPP policy is shown to be effective in balancing the trade-off between learning accuracy and resource constraints. The use of a surrogate penalty and the rectangular admission region allows for a tractable online implementation. The comparison to an oracle benchmark provides strong theoretical guarantees. However, the paper assumes that the sampling cost is known at each time step, which may not hold in practice. Future work could consider learning the cost distribution or incorporating uncertainty. Additionally, the retention horizon K is selected offline, which may not be optimal in non-stationary environments. Adaptive selection of K could further improve performance. Overall, the work advances the understanding of data management in streaming federated learning and provides a practical policy with theoretical guarantees.

CS practitioners and researchers

Opening member content…