A link between two wiring-closet switches in a clinic wing passes traffic, but throughput crawls and several devices report intermittent retransmissions, while every other link in the building performs normally. Interface statistics on that one link show a steadily climbing count of late collisions. What does this specific counter point to as the underlying cause?

Late collisions occur outside the normal collision window that half-duplex Ethernet expects, and the classic cause is a duplex mismatch: one interface is fixed at full duplex while its neighbor auto-negotiates or is fixed at half duplex, so the half-duplex side applies carrier-sense and collision-detection logic while the full-duplex side transmits whenever it wants, corrupting frames and forcing retransmissions that manifest as crawling throughput on an otherwise up link. An excessive cable length would degrade the signal more generally, producing CRC and runt errors from attenuation rather than the specific late-collision pattern tied to duplex disagreement. A full MAC address table causes the switch to flood unknown-destination frames out every port, which shows up as broader unnecessary traffic and CPU load, not a climbing late-collision counter on one specific link. A native VLAN mismatch on a trunk causes VLAN leakage or spanning-tree warnings and misdelivered tagged frames, which is a Layer 2 forwarding problem entirely separate from the collision counters. Caveat: late collisions can also appear on a genuinely half-duplex segment that is simply too long, so confirm both ends' negotiated duplex setting. A concrete check is comparing the duplex setting reported by each switch for that link — a mismatch there confirms the cause immediately.