IMPACT OF CLOCK DRIFT, EVENT LOSS AND NODE FAILURES ON DIAGNOSTICS IN IOT LOGGING

Distributed IoT logging subsystems operate on incomplete, out-of-order event streams, yet diagnostic methods are typically evaluated on complete ordered datasets. This paper empirically measures how clock divergence, event loss and partial node failures degrade diagnostic quality on a hardware testbed of five ESP32 nodes and a GPS/PPS-disciplined Raspberry Pi 3 aggregator. 448 runs of 300 s were performed under a combined design of one-factor sweeps and 3×3 grids. The inversion rate of the reconstructed order grows linearly with measured clock divergence (R² = 0.987) and matches the model (1 − 1/k)·d̄/T, with k nodes, mean divergence d̄ and run length T, to within 0.2 % at the top level; an order-sensitive detector's F1 drops from 0.96 to 0.72, leaving completeness unaffected. Conversely, the conventional F1 score (the harmonic mean of precision and recall) on delivered events is structurally insensitive to loss: it remains 1.000 both at 25 % loss and with three of five nodes failed, whereas system-level diagnostic completeness drops to 0.731 and 0.402 respectively; the share of detected incidents barely drops under loss (0.976), unlike failures (0.404). No interaction on completeness appeared in 3×3 grids. A common-mode offset leaves reconstructed order unchanged, so drift injection must be differential.