Skip to content

perf-notes-net.md

Latest commit

 

History

History
56 lines (42 loc) · 3.02 KB

perf-notes-net.md

File metadata and controls

56 lines (42 loc) · 3.02 KB
Raw

Lane B — network/interface optimizations (perf-trial-20260906-030748-net)

Status: COMPLETE. node --test: 34 pass / 0 fail. Note: dispatched subagent completed W7 (http pooling) + the mapLimit helper; the coordinator took over mid-campaign (2-subagent limit reduced to 1) and finished the rest.

W7 — Keep-alive HTTP connection pooling (commit bb616b3)

src/http.mjs: the secure path used Node's default global agent (keepAlive: false), so every Thunder/stargazer/cover/page call paid a full TLS handshake. Now one pooled agent per TLS mode: keepAlive: true, maxSockets: 8, maxFreeSockets: 4, keepAliveMsecs: 30_000, noDelay: true, agent-level timeout as an idle-socket backstop. Per-request timeout + req.destroy(...) handlers are untouched — the wedge failure mode the file header warns about stays closed.

  • Expected win: one RTT+handshake (~50–150 ms to OverDrive hosts) saved per reused connection. libby avail <100 ids> = 1 POST; archive typically makes 5–10 REST calls per loan plus every page/asset/cover.
  • Tradeoff: up to 4 idle sockets per host linger ≤30 s after the last request; CLI exit closes them. Bounded sockets (8) cap parallel-fanout blast radius.

mapLimit helper (commits 8e81da8, 6be8d36)

src/pool.mjs: bounded-concurrency ordered map, fail-fast. The dispatched agent's version declared failed but never set it — its doc comment promised "no further items are started" but free workers kept pulling. Fixed + added a test that pins the semantics (a freed worker must not start items after a rejection; the old test passed only by timing accident).

W8 — Concurrent audiobook part downloads (commit 6be8d36)

src/archive.mjs: spine parts download 3 at a time via mapLimit instead of strictly sequentially. Skip/resume byte-size logic, per-part logs, fail-fast, and partFiles order (spine order → manifest order) preserved. ~3× wall-clock on the download phase (network-bound; CDN is what the browser hammers with 6+ connections).

Tradeoff: 3 simultaneous connections per title to the CDN — deliberately low.

W9 — Concurrent page + asset fetching (commit 6be8d36)

src/archive-read.mjs: pages fetch+decode 4 at a time (a 300-page magazine was 300 serial RTTs — at ~80 ms RTT that's ~24 s of pure waiting), then assets 4 at a time. pageEntries stay in spine order for EPUB assembly; non-200 assets still log and get skipped; first page error rejects the loan (fail-fast, same as sequential abort).

  • Tradeoffs: in-flight bodies raise steady-state RAM by the concurrency factor (bounded, a few MB). Log lines now report decoded pages (completion) rather than starting pages, so numbers under concurrency are truthful. Cosmetic interleave.
  • Also removed: dead width variable in archive-read step 4.

What I verified

  • node --test: 34/34 (incl. new fail-fast pin test).
  • Static reasoning only for network wins (no live OverDrive calls from the trial); hub integration adds a local-TLS pattern bench to measure keep-alive + concurrency against a latency-injecting server.