gNMI, INT, and the metrics that matter
Two incidents, six months apart, on the same fabric.
First incident. A training job slowed by 18% starting at 02:14 UTC. The team noticed at 09:30 when checkpoint cadence reported the slowdown. They spent six hours hunting for the cause. Eventually traced it to a marginal transceiver on one spine port that was PFC-pausing intermittently. The counters showed it clearly, in retrospect.