运营商在 Fannel 接口上丢失和丢弃错误

During higher traffic times, we've noticed that the prometheus metric `node_network_transmit_carrier_total` increases on the `flannel0` interface. We're using the `xvlan` mode on k8s clusters running in AWS, managed by kops. We also observe unexpected network issues regarding node-to-node and node-to-external-service communication during periods where this counter increases (ie. `Library error: a socket error occurred`, or `read error on connection to (...)`).