FieldNotes
← All notes

HTTP/3 is not HTTP/2 over UDP

28 October 2026 · 3 min read

People describe HTTP/3 as “HTTP/2 with QUIC underneath”, which is true at the level of frame types and false at the level of behaviour. HTTP/2 multiplexes streams inside one TCP connection, so a single lost segment stalls every stream until the retransmission arrives. QUIC gives each stream its own delivery guarantee, so loss on one stream stops only that stream.

That difference shows up in places that have nothing to do with throughput:

Connection migration. A QUIC connection is identified by a connection ID rather than a four-tuple, so a laptop switching from Wi-Fi to a phone hotspot can keep the same connection. Over TCP that is a new connection and a new handshake, every time.

Middleboxes. TCP has forty years of middlebox behaviour baked around it. UDP does not, and some networks treat unknown UDP flows very differently from TCP 443 — which is why a site can feel fine over h2 and unusable over h3 on the same network.

Observability. There is no ss -ti for QUIC. You get packet captures and the connection's own qlog, if it emits one. Most of my debugging starts with “turn off HTTP/3 and see if it is still broken”.

The one config line that matters

add_header Alt-Svc 'h3=":443"; ma=86400' always;

That header is how a browser learns the site speaks h3; without it, clients stay on h2 indefinitely. If you remove it, browsers that already learned the route will keep trying h3 for up to the ma lifetime, which is a good reason to keep it short when you are experimenting.

None of this makes h3 the right default for a small text site. The win is real on lossy mobile links and mostly invisible on a fibre connection to a server in the same city.