Large-scale distributed training is the driver of modern artificial intelligence, but its performance is critically vulnerable to “stragglers”, i.e., workers that slow down the entire process, leading to immense resource waste in industrial settings. Existing profiling tools focus on single-node analysis and lack a global perspective to identify systemic bottlenecks in distributed settings. In addition, recent emerging performance analysis tools can handle stragglers but lack diagnostic insights that engineers need to determine root causes. We introduce KAIR, a production-grade observability system that fills this critical diagnostic gap. KAIR is built on two core principles: (1) a low-overhead, scalable architecture to collect and aggregate fine-grained traces from thousands of nodes and (2) a novel hierarchical analysis engine that moves from statistical anomaly detection to causal inference. KAIR moves beyond simple timeline visualization by introducing a suite of techniques, including Kolmogorov-Smirnov statistics to identify divergent/unhealthy nodes and z-score analysis to pinpoint anomalous operators. This allows KAIR to pinpoint not only which worker is a straggler, but the specific root cause responsible for the delay. Evaluations were performed in Kuaishou production clusters. The experiment results indicate KAIR is highly effective in identifying latent performance bottlenecks overlooked by conventional tools. It performs negligible overhead, scales to large clusters, and provides actionable insights that have significantly reduced computational waste and engineering efforts.