Possibly you'd have more luck on a network where your client can allow incoming UDP connections on the Tailscale port, and so the exit node would be able to establish a direct connection.
But for a Tailscale peer I have running on AWS ECS, I can open the UDP port there, so a direct connection always happens regardless of what sort of network my Tailscale clients are on. I don't know if there's any Fly equivalent to get a direct connection to a UDP port.
I just tested it again and the connections are made directly (after the first 2,3 packages go via DERP):
tailscale ping fly-ams
pong from fly-ams (100.96.123.32) via DERP(ams) in 15ms
pong from fly-ams (100.96.123.32) via [2604:1380:4601:d605:0:6c3b:eed5:1]:41641 in 12ms
tailscale status
100.96.123.32 fly-ams patte@ linux active; offers exit node; direct [2604:1380:4601:d605:0:6c3b:eed5:1]:41641
100.101.54.36 fly-hkg patte@ linux active; offers exit node; direct [2605:4c40:95:4eed:0:40f0:67b1:1]:41641
[1]: https://github.com/patte/fly-tailscale-exit/blob/main/fly.to...
[2]: https://github.com/patte/fly-tailscale-exit/blob/main/start....Fly.io or not, this was an issue I always ran into with Tailscale.
They talk a big game about NAT punching, and using various UDP shenanigans to get around P2P connection formation issues, but at the end of the day, most of my connections were via DERP, even with fairly trivial firewall configurations.
The downside is that my traffic never went direct; it was always relayed via a Tailscale DERP node, as Fly.io machines were only accessible via anycast, and so a direct connection from Tailscale on my machine to the exit node on Fly.io couldn't be established.
So performance wasn't as great (and I felt bad about using up Tailscale's DERP bandwidth, as a free user).