HTTP/3 is not HTTP over a faster pipe; it is HTTP over a rebuilt transport. QUIC moved the machinery TCP kept in the kernel, streams, reliability, congestion, into userspace over UDP, and several operational behaviors changed in ways delivery teams should understand rather than trust.
Streams without shared fate
QUIC multiplexes independent streams inside one connection, each with its own flow control and loss recovery. A lost packet stalls only the streams whose data it carried, unlike TCP, where one hole halts every multiplexed byte behind it. This is the head-of-line fix, and it matters most exactly where loss lives: cellular, wifi at conference-center density, oversubscribed international transit.
Connection IDs and migration
QUIC connections are identified by connection IDs, not the IP/port four-tuple. When a phone hops from wifi to cellular, the connection survives the address change and continues after a round trip of validation, no re-handshake, no session teardown. For long-lived transfers and live streams on mobile audiences, migration converts a network handoff from a visible stall into a non-event, and it is invisible in every lab benchmark run from a machine that never moves.
One more internals note with buying implications: because QUIC lives in userspace, CDN vendors iterate transport behavior at software speed, pacing strategies, ACK policies, recovery heuristics ship in releases, not kernel upgrades. This is why HTTP/3 performance between providers diverges more than HTTP/2 ever did, and why the divergence changes quarter to quarter. The practical consequence for buyers: protocol benchmarks age quickly, and a provider’s transport-engineering culture, do they publish, do they measure, do they iterate, predicts future performance better than any single snapshot. Ask when their QUIC stack last changed and what moved; the quality of the answer is the answer.
Loss recovery and the encryption trade
QUIC’s acknowledgment and recovery machinery is richer than TCP’s, with better reordering tolerance and per-packet timing. Nearly everything is encrypted, including transport metadata that middleboxes historically inspected, which ends a whole category of network meddling and also blinds some enterprise proxies and legacy firewalls; graceful HTTP/2 fallback therefore remains mandatory. Server-side, all of this executes in userspace, so implementation quality varies by vendor far more than TCP stacks ever did.
In practice
Operational checklist: confirm UDP 443 flows freely along your paths, verify fallback behavior by blocking it deliberately in a test, read your provider’s cohort analytics for H3 share and per-protocol percentiles, and pay attention to mobile-cohort tail latency, where streams, migration and modern recovery collectively earn their keep. If your provider cannot show per-protocol splits, that observability gap is itself a finding.
Per-protocol, per-cohort benchmarking is standard in our assessments. HTTP/3 is where the interesting deltas live.