Answers are grounded in this paper and cite their sources.
IoT telemetry pipelines in distributed web systems forward large volumes of mostly nominal sensor data to the cloud. This paper studies a three-tier pipeline in which edge gateways run a sliding-window z-score (SW-Z) filter, Apache Kafka decouples ingestion from analytics, and Kubernetes microservices are controlled by an energy-aware scheduler. Every 15 s, the scheduler solves, by exhaustive enumeration over evenly spread placements, a replica-selection and placement problem whose power model uses a published server load-power curve, node power-down and the per-poll cost of Kafka consumers; with identical, always powered nodes, a linear power curve and only work-conserving services the energy term would be constant. The filter is evaluated offline on NASA C-MAPSS data, and the cloud tier with a trace-driven simulation calibrated with measured per-message costs (10 seeds per point). At threshold 3, SW-Z reduced upstream payload by 57.4% (95% CI 56.2% to 58.5%) on FD001 and by 61.0% on FD003, and raised an alarm for 94.2% of injected 5σ anomalies, but with precision 0.239. In the simulation, under constant load below uplink saturation, mean P99 end-to-end latency stayed between 25.4 and 30.2 ms in every configuration except a lag-based autoscaler, whose consumer-group rebalances raised it to between 43 and 388 ms on average; filtering raised the computed uplink capacity from about 55,600 to 128,000 records/s. At 50,000 records/s the scheduler reduced simulated mean cluster power from 402.8 W under CPU-based autoscaling to 218.0 W, and to 199.1 W with filtering, mainly by avoiding the scale-out of polling consumers.
No takes yet. Share an insight, caveat, or question.
Ileana Marian (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: