Personally, I think it's a large packet loss burst at the time when the initial client pings are being done, so the pings don't get done, or the client stats don't get sent up from client -> server.
Make the whole client ping more resilient, so the server can detect, wait, I didn't actually get any client ping stats sent up to me yet, and the backend sends down a new set of client relays to ping and it tries again.