Starting around 09:00 UTC, one of the nodes where a VPN service pod was running encountered memory pressure causing VPN connections to drop. The amount of reconnecting devices started overloading the other VPN service pods. This made the VPN connections to intermittently drop as the connections were being redistributed to recovering VPN service pods. The devices were still able to communicate with the backend as soon as a VPN tunnel was established.
We increased the cluster capacity and added additional VPN service pods to handle surges of new VPN connections.