Answers are grounded in this paper and cite their sources.
Continuous telemetry from heterogeneous Internet of Things (IoT) devices strains cloud-centric streaming: arrival rates vary, uplink capacity is non-stationary, and many applications require a hard end-to-end latency bound. Priority queues protect a latency class and operator placement enforces a response-time constraint, but neither jointly tunes the micro-batch window, the compression level, and the offload fraction. This paper formulates that joint decision for periodic, event-driven, and bursty streams sharing one edge uplink. An online controller predicts mean per-record delay from batching, edge service, wide-area transmission, and uplink congestion, and selects a window, a codec, and an offload fraction between a fidelity floor and a safety-scaled latency bound. In a time-stepped emulation of 24 synthetic streams offering 11.29 MB/s to a 12.0 MB/s uplink, the controller reduces the record-weighted deadline-miss ratio from 0.166 (cloud-only) and 0.223 (priority queues) to 0.020 and cuts cloud ingress from 11.27 MB/s to 2.49 MB/s, while retaining a mean fidelity of 0.92. At 1.5 and 2.0 times the nominal load, both forwarding baselines saturate the uplink (miss ratio above 0.99), whereas the controller holds at 0.032 and 0.050. A fixed edge configuration (chosen in advance, 200 ms window, medium codec) achieves more reliable deadline adherence, missing at most 0.2% of records at any load with similar ingress, but at 43% higher modelled edge energy consumption. An ablation shows that the controller's residual misses derive from its choice of window: because the mean-value predictor only accounts for half a window of batching delay, it chooses windows for which the oldest records exceed the bound, suggesting a worst-case batching term as a direct fix.
No takes yet. Share an insight, caveat, or question.
N et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: