Skip to content

CONCEPT Cited by 1 source

QUIC transport

Definition

QUIC is a user-space, UDP-based transport protocol standardised by the IETF as RFC 9000 (May 2021). Unlike TCP — a kernel-space, byte-stream transport designed in the 1970s — QUIC is built around the protocols-cooperation principle: it fuses transport-layer, cryptographic, and stream-multiplexing concerns into a single user-space layer running over UDP.

Zalando's 2024-06 post frames QUIC as the "best technical solution" to the TCP-era protocol design inefficiencies that HTTP/1.1 and HTTP/2 inherit (Source: sources/2024-06-17-zalando-next-level-customer-experience-with-http3-traffic-engineering).

Key properties

  1. User-space flow + congestion control over UDP. QUIC gives up kernel TCP's ubiquity for user-space mutability: the congestion-control algorithm, stream scheduler, and flow controller live in a library the application ships (see concepts/user-space-congestion-control).
  2. Fused TLS 1.3 + transport handshake. One handshake negotiates both cryptographic and transport parameters, cutting cold-start RTTs compared to TLS-over-TCP (see concepts/quic-tls-fused-handshake).
  3. Stream multiplexing with independent flow control. QUIC multiplexes many application streams over a single UDP flow. Each stream has its own flow-control window — loss on one stream does not block the others at the transport layer (no transport-level HoL blocking, unlike HTTP/2-over-TCP where a lost segment stalls all streams).
  4. Connection identity decoupled from 4-tuple. QUIC connections are identified by a connection ID, not by (src-ip, src-port, dst-ip, dst-port). Clients can change network paths — Wi-Fi ↔ cellular — and keep the QUIC connection alive (see concepts/quic-connection-migration).
  5. Encryption mandatory at every packet. QUIC has no plaintext mode; TLS 1.3 encrypts the payload and header- protects packet numbers. Middleboxes cannot rewrite QUIC fields the way they could for TCP.
  6. Packetisation + fragmentation in-protocol. QUIC controls how payload is framed into UDP datagrams, enabling alignment with the congestion window and explicit MTU discovery (subject to the IPv6 1280-byte floor).

Why "user-space" is the design pivot

Zalando's post makes the hardware vs software ownership shift explicit:

"Historically, congestion control was owned by 'hardware' companies — those who developed networking equipment and operating systems. QUIC shifts the ownership, because of user-space implementation, towards 'software' companies — those who own Web-browsers." (Source: sources/2024-06-17-zalando-next-level-customer-experience-with-http3-traffic-engineering)

This shift is what enables BBR's rapid deployment, Deep- Reinforcement-Learning CC research (Aurora, Eagle, Orca, PQB), and Cloudflare / Google / Meta running differentiated CC in production without kernel patches. The price is per-packet kernel-crossing + user-space copy overhead.

Operational costs QUIC inherits

The post enumerates four open costs — all real engineering problems rather than protocol bugs:

  • Datagram-level HoL at the packetisation layer. Streams are multiplexed into coalesced UDP datagrams; if a datagram is lost, all streams it carries stall. Application-layer traffic prioritisation is needed to mitigate.
  • Memory pressure from out-of-order buffering + user-space copies. Kernel TCP amortises reordering in ring buffers; QUIC libraries must manage reordering in user-space, often copying from kernel to user memory.
  • UDP middlebox hostility. ISPs apply different routing / QoS / AQM policies to UDP than to TCP (see concepts/udp-middlebox-hostility). Production rollouts often require fallback to HTTP/2-over-TCP for some clients.
  • 1280-byte IPv6 MTU floor. QUIC datagrams cannot be smaller; networks operating below that (non-standard IPv4, some radio channels) cause fragmentation. Path MTU Discovery is required for larger datagrams.

Historical context: why prior attempts failed

Earlier transport improvements (SST, SCTP, MP-TCP, uIP, lwIP) stayed confined to the transport layer without considering the end-to-end Web perspective. QUIC's success — where those failed — is Zalando's framing: "built over protocols cooperation principles rather than a strict OSI layering." The decision to let transport and application layers cooperate (stream scheduler sees what streams are urgent, TLS integrates with transport state) is the architectural pivot that made adoption economic.

Adoption anchor

At the time of Zalando's 2024-06 post, w3techs measured QUIC at 8.0 % of websites worldwide. HTTP/3 (QUIC's application-layer partner) was at 29.8 %. Akamai supported QUIC since 2016; AWS Network Load Balancer supports QUIC for UDP load-balancing; CloudFront terminates HTTP/3 at AWS Edge Locations. Zalando uses this stack for its own HTTP/3 media delivery.

Seen in

Last updated · 501 distilled / 1,218 read