DeadOrSlow: silence-aware client failure detection for semi-synchronous federated learning (code, raw logs and design log)

Abstract

Reproducibility artifact for the manuscript "Dead, Hung, or Just Unreachable? Silence-Aware Client Failure Detection for Semi-Synchronous Federated Learning" (Z. Kutlu and R. K. Ağca). Contents: the detector code shared by the online FL server and the offline trace replay; the FL server and Raspberry Pi client; the fault and network-disruption schedulers; raw event logs, round logs and fault logs of every campaign on a four-node Raspberry Pi cluster (detector-level campaigns, closed-loop campaigns with 4, 8 and 20 clients, and a negative control in which training stalls while heartbeats continue); the dated design log (in Turkish) with every pre-registered hypothesis; and the scripts that regenerate every table and figure of the manuscript. See README.md inside the archive. Version 1.0.1: before submission, an independent audit found that the fault injector of the 8- and 20-client runs applied the events of each board sequentially, so some scheduled faults started late or not at all. The scale analyses now take the ground truth from the boards' fault logs (rq2/realized_faults.py). Round counts, false exclusions and all pre-registered decisions are unchanged; round-extension medians and the number of hangs near outages changed. New scripts: realized_faults.py, scale_realized_report.py, scale_fe_detail.py, train_time.py; DESIGN.md section 18. Raw data are identical to version 1.0.0.