Tier 1 default-transport validation
Verdict. Correctness-backed Tier 1 default selection; not an overall performance win. Ordinary-route throughput vs old Tier 1 ranges -5.1% to +3.7% in median with overlapping spreads; POST c512 and known-length stream c512 p99 medians regress. No leaderboard update.
Errata and caveats
- Socket-option erratum. Historical
tcp_nodelay=off/ override0labels did not prove TCP_NODELAY was disabled: the old client skipped its override while NIO enabled TCP_NODELAY by default, and its explicit1used the wrong option level. Raw measurements are preserved; rebaseline with verified socket values (report header; the companion correction report is not in this site's data pack). - Client event-loop distribution. The benchmark client requests four threads but uses two loops; this harness limitation is held constant across the servers in this matrix (report, Network experiment).
- VM scope. Docker Desktop Linux VM results do not establish bare-metal results; bare-metal Linux with a separate generator and an independent harness remains necessary before generalizing (report, Tail-latency limits).
Sources
- Full report (rendered)
- Full report (verbatim Markdown)
tier1-default-2026-10-18.json— 432 per-round samples, SHA-25684d034b170aa7d374f3970dee7578c244210a900d594e7f77cd93003653fe70d
WhiteThroated