Hudson River Trading’s October 12 engineering account follows a research-network bottleneck between New York data and Chicago compute. Its application logs could not explain all the traffic. Kernel-level tracing, joined to workflow records and switch measurements, exposed repeated remote file reads: thousands of jobs were reaching across the network instead of benefiting from one local copy. The file cache made those reads fast enough to overwhelm the link. The diagrams simplify a real incident; this is not a trading-latency benchmark. The useful distinction is between measuring a service and attributing the work that caused it. HRT engineering