A 911 records lake fabric must survive a leaf failure without stranding half the GPUs. What network design property should evaluation prioritize?

A 911 records lake fabric that must survive a leaf failure without stranding half the GPUs should prioritize redundancy so alternate leaf and uplink paths keep GPU servers reachable after a failure. Dual-homing and spine diversity shrink blast radius: when one leaf fails, public-safety AI workloads continue on alternate paths instead of losing a large GPU cohort and the analytics that depend on it. Evaluation that skips this property accepts an outage shape incompatible with emergency records workloads. Multi-chassis link aggregation to one leaf to keep sessions up during an uplink outage helps survive a link loss, but it still anchors GPU servers to a single leaf and does not keep them reachable when that leaf fails. Redundant spine supervisors with stateful failover to retain routes after a control failure protect the control plane on a spine, yet they cannot provide an alternate leaf or uplink path to keep GPU servers reachable when a leaf fails. Fast link detection with sub-second timers to reroute flows around a failed spine link speeds convergence around a spine-link fault, but it offers no alternate leaf attachment or path diversity to protect GPU cohorts when the leaf itself fails. Exam caveat: AI network evaluation includes redundancy—leaf and uplink diversity so one device failure does not strand GPU cohorts—especially for public-safety workloads. Operational check: in design review, prove dual-homing of GPU servers, verify spine diversity, simulate leaf failure, and confirm the 911 records lake training and inference endpoints remain reachable without a full halt.