HTTP internals compared

Measured on one machine on one day, by standing each server up and probing it — not read from documentation. Reproduce with bash bench/probe-runtimes.sh.

A silent connection — one that completes the TCP handshake and then says nothing — is the cheapest hold on a server there is: one syscall to the peer, no state to keep. Node bounds it at 88s, which is a 60s headersTimeout polled on a 30s interval. Deno does not bound it at all, despite being built on the same HTTP implementation we are; nor does it bound an idle keep-alive connection.

The HTTP/2 rows are read off the wire from each server's SETTINGS frame, which is what it tells every client its limits are. Node and Bun advertise unlimited concurrent streams; the runtimes that cap are the two on hyper.

None of this makes one runtime better than another — Node's 5s keep-alive is more aggressive than ours, and its 88s header bound is looser. It is here so that a number in this documentation can be checked rather than believed.

See also

• runtime:http API reference — signatures, options, errors

• HTTP server guide — how to build one

• Internals: sockets — runtime:net, which this does not share code with

• Internals: the fetch client — the outbound direction

• Benchmarks — throughput, including HTTP/1.1 vs HTTP/2

• vs Node · Bun · Deno — capability-level comparison