kabirx

Peer relays are a bit different from our previously available Custom DERP servers. While the custom DERPs do relay traffic, they also require a bunch of configuration and management for their other jobs and they open up availability concerns that are pretty tough for our average customer.

Conversely Peer Relays are built on top of the shoulders of DERP. For example, they don't need to do peer discovery set connections up end to end - instead connections are brokered via our DERP fleet and then in a sense "upgraded" to an available Peer Relay or Direct connection. Because of that they're super lightweight and much easier to deploy + manage. And, they scale horizontally so you can deploy many peer relays across your network, and they're resilient to downtime (we'll just fall back to DERP).

(Tailscalar here) We're taking this kind of outage very seriously. In particular this outage meant newly connected devices couldn't reliably reach our control plane and couldn't get the latest network state. IMO that's not okay.

One of Tailscale's fundamental promises is that we want to try as much as possible to get our control plane and infrastructure as out of the way of your connectivity paths, while still using our infra to "assist" when there's connectivity issues (like difficult to traverse NAT), and maintain trust across the network, and keep everything up to date.

It's a tough balance and this year we're dedicating resources to making sure even small blips in our control plane don't mean temporary losses of connectivity across even your newly woken up devices. In particular we're taking a multi-pronged approach, right now. We're working in parallel to increase client tolerance of control outages (in response to cracks shown in this incident) and have an ongoing effort to make the control plane more resilient and available.

One limitation of custom DERP is that across tailnets, they don't share the same DERP maps and don't have access to each others' DERPs.

With Tailscale Peer Relays, the available relay bindings can be seen by the devices on either side of a connection; as such it should work out of the box with a sharing relationship between tailnets.