Every few months a QUIC rollout gets blamed for a problem that turns out to be an MTU black hole somewhere in the path. The pattern is always the same: TCP works, QUIC stalls, and the packet captures show large datagrams leaving but nothing coming back.
TCP gets segmented by the kernel and can rely on retransmission with a shrinking MSS. QUIC has no such negotiation — it relies on DPLPMTUD, and some middleboxes drop the ICMP "fragmentation needed" messages that would let it converge. The result is a handshake that completes and a transfer that never starts.
Path MTU on the tunnel interface, then the clamp on the router. On most Linux boxes a mss clamp plus a sane advmss on the tunnel gets you out of trouble; the important part is testing from the actual client network, not from the datacentre.
Second on the list: any device doing NAT that rewrites the outer header but not the inner one. That one cost me an afternoon.