Skip to content

HTTP proxy/LB bench regression fix (wave 8 revert + argv routing)

Summary

Fixes tier-5 proxy_loopback / LB wrk_parse_fail_li by reverting broken wave-8 httpd_proxy_resp_finish_headers filter relay and routing CSV upstream ports to the Li epoll proxy loop (not static httpd_serve_port_root).

Agent continuation

  1. Read: runtime/li_rt_net.c (httpd_proxy_snap_reset, httpd_proxy_snap_begin_recording_if_get); packages/li-net-httpd/src/lib.li (httpd_upstream_proxy_epoll_loop, httpd_run_from_argv).
  2. Run: build li-httpd; from benchmarks BENCH_PROXY_ORACLES=nginx,li ./scripts/run-tier5-http-bench.sh; check proxy_loopback / lb_* rows in docs/ecosystem/http-server-rps-matrix.md.
  3. Then: re-apply wave-8 x-internal header strip with a safe filter (do not break httpd_proxy_resp_finish_headers body relay).
  4. Blocked on: none for loopback proxy RPS.

Changed

  • runtime/li_rt_net.c — snap race guard (g_proxy_snap_recording); httpd_proxy_snap_reset on upstream set; pre-wave-8 proxy finish path (no broken header-filter relay).
  • packages/li-net-httpd/src/lib.li — httpd_upstream_proxy_epoll_loop; CSV argv → proxy loop (not static serve).

Not changed

  • g_strip_internal_headers wave-8 feature (deferred safe re-land).
  • li-tls terminate / HTTPS listen.
  • benchmarks harness (updated in benchmarks PR).

Breaking

N/A.

Security

N/A — bench fix only.

Performance

Loopback proxy (representative): proxy_loopback li ~130k+ RPS vs nginx ~78k on same host after fix.

Downstream

  • benchmarks: refresh matrix; merge supplemental http_tier5.csv only when li RPS > 0.