Networking & Protocols
HTTP versions: multiple-choice review
Crux Multiple-choice synthesis across the HTTP-versions unit — multiplexing, layered HOL blocking, QUIC stream isolation, header compression, and production protocol selection.
Your altitude — climbing toward senior
ZeroJuniorMiddleSenior
You are at senior altitude — in orbitSix questions that cut across the whole unit. Each mirrors a protocol-selection or diagnosis call you make in production — not a definition to recite, but a tradeoff to weigh under real network conditions.
Goal
Confirm you can connect multiplexing, the two layers of head-of-line blocking, QUIC’s stream isolation, header-compression ordering, and the operational signals that decide which HTTP version actually wins.
Quiz
Completed
An API does 50 concurrent sub-requests per page. On a fibre path (loss under 0.1%) HTTP/2 is fast; on a 1.5%-loss cellular path it is measurably slower than HTTP/1.1 with 6 connections. Why?
Heads-up The 9-byte HTTP/2 frame header is negligible. The regression is TCP-layer HOL blocking: one lost packet stalls all multiplexed streams on the single connection.
Heads-up HPACK does not corrupt under loss; TCP redelivers in order. The issue is that the shared TCP byte stream stalls every stream while the lost segment is retransmitted.
Heads-up Carriers do not throttle by HTTP version. The cause is structural: a single TCP connection has one loss-recovery domain, so loss stalls every stream at once.
Quiz
Completed
HTTP/2 multiplexes independent streams, yet it still suffers head-of-line blocking. At which layer, and why does HTTP/3 escape it?
Heads-up HTTP/2 streams are independent at the HTTP layer; interleaving is free. The remaining HOL blocking is purely transport-layer, from TCP's ordered byte stream.
Heads-up Stream count is not the point. QUIC gives each stream its own loss recovery, so a lost packet stalls only the streams whose data it carried — TCP cannot.
Heads-up QUIC still retransmits lost stream data reliably. It isolates loss per stream rather than abandoning reliability.
Quiz
Completed
Why did HTTP/3 replace HPACK with QPACK, and what would happen if it had kept HPACK over QUIC?
Heads-up Compression ratio and speed are comparable. The change is about ordering: HPACK's shared ordered table defeats QUIC's stream independence; QPACK preserves it.
Heads-up Both HPACK and QPACK already defend against header compression-oracle attacks. The reason for QPACK is the stream-independence ordering problem on QUIC.
Heads-up QPACK actually keeps a static table (99 entries). The real driver is decoupling dynamic-table updates so one stream's headers never block another's.
Quiz
Completed
A team plans to ship HTTP/2 Server Push to preload critical CSS. What should a senior reviewer say?
Heads-up Push looked good on cold-cache benchmarks but lost in production: blind re-pushing of cached assets, no help on warm loads, and middlebox breakage. It is deprecated.
Heads-up HTTP/3 does not rehabilitate push; the cache-visibility problem is identical. The modern replacement on any version is 103 Early Hints plus preload.
Heads-up Compression is unrelated. The defect is operational — the server guessing what the client cached — not the encoding of pushed bytes.
Quiz
Completed
An operator turns on HTTP/3 at the edge. Overnight the ALPN dashboard shows h3 share collapsing from 21% to 1% with no application errors. Most likely cause and right first move?
Heads-up An expired cert would break HTTP/2 over TCP too and surface loud errors. A silent h3-only collapse with h2 absorbing traffic points to UDP/443 being blocked on a path.
Heads-up QPACK table exhaustion degrades compression, not protocol selection. A full ALPN collapse to h2 is a UDP-reachability problem, not a header-compression one.
Heads-up Browsers do not randomly rotate versions; they pick h3 when Alt-Svc or SVCB advertises it and UDP works. A 20x drop is a network event, not noise.
Quiz
Completed
A real-time multiplayer backend serves 100k mostly-cellular connections. Position updates are useless if they arrive late but a missed one is harmless. Which stack fits best, and why?
Heads-up RFC 8441 saves connection setup but rides TCP: one lost packet on lossy cellular stalls every game-state stream. It also lacks unreliable datagrams for fast position updates.
Heads-up Each WebSocket needs its own TCP connection paying a cold handshake on every mobile reconnect, with no loss isolation and no unreliable channel — the worst fit for lossy real-time.
Heads-up Raw UDP re-implements congestion control, encryption, NAT and mobile-firewall traversal, and anti-replay — all of which QUIC already provides. Only justified for extreme niche cases.
Recap
The unit’s through-line is one progression: HTTP/1.1 parallelises with many connections, HTTP/2 multiplexes streams on one TCP connection (removing application-layer HOL blocking but inheriting transport-layer HOL blocking), and HTTP/3 moves to QUIC for per-stream loss recovery, QPACK header compression, and connection migration. The production lens is constant — match the version to the network (HTTP/3 on lossy and mobile, HTTP/2 on reliable east-west), watch ALPN and h3-fallback-by-ASN as your health signal, and prefer 103 Early Hints and RFC 9218 priorities over the dead Server Push.