We have monitored packet losses in TTN very thorowly during the last time (see dataloss-in-the-backend) and the backend currently eats about 25% of our packets.
Regarding the original SPF things are really a bit funny. Semtech has implemented an uplink acknowledge, so the gateway really knows, if a packet was transmitted or not. It just does not use it. Kerlink has implemented a new forwarder they called common packet forwarder (CPF), that can retransfer unacknowledged packets. That helps, if a single packet was lost over UDP.
We found that loss rates are still high. If a packet is lost between gateway and backend, it ist not possible to distinguish if loss was caused by UDP or if the backend was just not listening for some times. So it may be an UDP issue, but also maybe caused by some other part of the system.