Train v: web speed at 100k notes, delete race (#327, #448)

kayg commented

2026-10-17 07:06:33 +00:00

Owner

Train v ran the full suite per project (11 projects), plus fmt, clippy, workspace tests and the build. fmt, clippy and the workspace tests are all green.

Passed: chromium-1440, throttled-390, throttled-1440, flow, passkeys, prefs-chromium, prefs-webkit and contrast.
First-run failures, all green on a 3x rerun:
- missing-links:379 on chromium-390: 54 of 54;
- tooltips on webkit-390: 9 of 9;
- tree:176 on webkit-1440: 21 of 21.

Contents:

#518: web speed. The 100k cold TBT drops from 22.5 s to about 0.4 s on an idle host, with the stale-list guard kept;
#519: delete waits for the history writer, so a link read right after a delete never says New.

Merging.

Replaces #518 and #519.

🤖 Generated with Claude Code

Train v ran the full suite per project (11 projects), plus fmt, clippy, workspace tests and the build. fmt, clippy and the workspace tests are all green. - **Passed:** chromium-1440, throttled-390, throttled-1440, flow, passkeys, prefs-chromium, prefs-webkit and contrast. - **First-run failures, all green on a 3x rerun:** - `missing-links:379` on chromium-390: 54 of 54; - `tooltips` on webkit-390: 9 of 9; - `tree:176` on webkit-1440: 21 of 21. Contents: - #518: web speed. The 100k cold TBT drops from 22.5 s to about 0.4 s on an idle host, with the stale-list guard kept; - #519: delete waits for the history writer, so a link read right after a delete never says New. Merging. Replaces #518 and #519. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

kayg added 18 commits

2026-10-17 07:06:34 +00:00

Perf: web client harness and budgets on big Homes (#327 ) cc653ed74c

perf/web.sh measures the built app as production serves it on 10k and 100k
fixture Homes, Chromium (CPU 4x) and WebKit: bundle sizes, boot (TBT as
web.dev defines it, every long task, first paint, time to interactive),
keystroke to paint, the blaze test (#271), scrolling, memory, layout shift
and main-thread CPU. web.* budgets in perf/budgets.json; CI gates the
bundle sizes. docs/perf/web.md.

Replaces #360.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018Yyy8fQGxPLkLa8v8NXz7T

Editor: no whole-text copies per key; a preview goes live in place (#327 ) fadf4c0c71

- frontmatterEnd walks lines without copying and remembers per document.
- The note metadata keeps `base` under its own key, written only when it
  changed (older records still read).
- A preview whose text matches the live document swaps its binding in
  place (a compartment) instead of building and parsing a new state.
- Clouds worked out in slices (lib/slices.ts), date labels morph only once
  tapped, the offline merge code loads only when a merge runs.

Replaces #402 and part of #454.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018Yyy8fQGxPLkLa8v8NXz7T

Web: notes list index, windowed rows, fill pages read in a worker, first note over HTTP (#327 ) abfff8b066

- lib/notes-index.ts: notes newest first by binary search, tag counts,
  file to id, the Journal's order; views are told once per batch.
- groups() slices ordered lists by binary search; NoteList works out its
  visible rows from group sizes (lib/list-window.ts). Nothing walks every
  note on the main thread.
- The background fill (#422, pages of 2000) is fetched and parsed in a
  worker (lib/notes-fill.worker.ts) and handed over 250 notes at a time;
  without a worker, pages of 500 the plain way.
- The first note's text comes over plain HTTP (notes.peek) when nothing is
  remembered, instead of waiting for its room; peeks come 12 at a time.
- Typing: properties re-read only when an edit reaches them, the bar's
  block list replaced only when it held something.
- Theme colour once per frame; the Settings bar builds at the next idle.

Replaces #394, #407 and the rest of #454.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018Yyy8fQGxPLkLa8v8NXz7T

Perf: wait for the whole list by the app's own mark, not request silence (#327 ) 6140d51abd

Fill pages are fetched by a worker now, whose requests the page doesn't
always see: the app marks html[data-listed] when the whole list is in, and
the harness waits for it (builds without the fill worker: request silence).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018Yyy8fQGxPLkLa8v8NXz7T

Perf: measure background list fill (#327 ) 5dc1f68aaf

Perf: cap worker fill chunks and measure overlap (#327 ) 3f1ce3b278

Perf: halve note fill chunks ac8ed7e679

Perf: hand off fill notes singly 9190c6b6d4

Perf: backpressure note fill messages 4736a0e21d

Perf: acknowledge note fill batches 759d84f60b

Perf: account fill tasks by handler work aac3890663

Perf: slice notes fill work on the page 6644051128

Links: ignore deleted notes in late list reads 513951e948

Test stale note lists deterministically

CI / server (pull_request) Successful in 26s

Details