Skip to content

Performance and measurement ​

TechPedia targets shared-CPU VPSs, slow networks and low-cost phones. Article text must be readable before JavaScript. Optional tools stay out of ordinary reading dependencies.

Low-end regression profile ​

Use 360×640 at DPR 2, 400ms RTT, 400kbps download/upload, Chromium 6× CPU throttle, cold cache and reduced motion. This is not an exact device simulation.

MetricBudget
FCP≤8s
Local uncompressed same-origin JavaScript≤240 KiB
CLS≤0.1
Longest task under 6× throttle≤1s
Horizontal overflow≤1px
Readable body>20 characters

After building and serving an isolated article:

sh
node tests/performance/low-end-audit.mjs 'http://127.0.0.1:18090/en/wiki/Example'
node tests/performance/low-end-audit.mjs --desktop 'http://127.0.0.1:18090/en/wiki/Example'

Set CHROMIUM_PATH if needed. The CDP script reports JSON. Test only your own service and fixtures.

Boundaries ​

Preserve SSR DOM through Vue mounting and reading/discussion changes. Public reads do not wait for auth. Protected routes check permission before admin imports. Cancel stale requests and ignore late generations on article/history changes. Read headings from actual DOM; resolve/decorate links in yielding bounded batches. Defer offscreen layout without removing source, anchors or print content. Editor/admin/graph/QR/gallery code loads on demand. Bound caches and cold-render work separately. Use responsive media variants and lazy/async image decoding.

Historical evidence, not an SLA ​

A 2026-07-18 local 7,266-character/80-heading fixture measured FCP 5,340ms, JavaScript 166 KiB, CLS 0.004 and longest task 926ms under the low-end profile. A later production article with ~30,000 DOM nodes failed task budgets: DOM/style cost, not only transfer size, mattered.

October 2026 warm microbenchmarks on Apple M4/macOS arm64, CGO disabled:

FixtureBaseline → improvedCumulative allocation
8 summaries / 500 revisions76.634 → 0.063ms~108 MiB → 13.6 KiB
8 search results / 500 matches5.006 → 0.095ms~8.14 MiB → 16.2 KiB
Warm render, 276KB source23.417 → 0.182ms~24.6 MiB → 16B

Allocation is not peak RSS. Short-source latency did not improve in every case. Search still scans candidates; it is not a persistent full-text inverted index.

sh
CGO_ENABLED=0 ./tests/go.sh test ./internal/storage ./internal/techmark -run '^$' \
  -bench '^Benchmark(RevisionSummaryRead|WarmDocumentSearch|WarmArticleRender)$' \
  -benchmem -benchtime=200ms -count=3

A separate synthetic 239,019-character/20,835-node article measured FCP 7,468ms and longest task 481ms after offscreen layout work. It was not the same production article and does not establish deployed performance.

Graph scaling ​

The public graph defaults to all document metadata, not a 1,500-node slice. It avoids the bot analyzer's dense PageRank. Browser Worker physics uses recycled numeric buffers and decreases update frequency as work grows. Full JSON transfer and GPU memory still scale with size; unlimited nodes/FPS are not promised.

Historical synthetic server construction (excluding DB, JSON and HTTP):

Documents / directed linksTimeCumulative allocation
2,000 / 4,0001.95ms2.19 MiB
10,000 / 20,00010.05ms12.15 MiB
50,000 / 100,00060.75ms61.06 MiB

Node physics medians were 6.30/37.57/229.96ms for 2k/10k/50k nodes, not full browser frames. A synthetic 10,001-document fixture was checked at desktop and 390px widths. Those historical checks are not current real-phone verification.

sh
./tests/go.sh test ./internal/server -run '^$' -bench '^BenchmarkGraphSnapshot$' -benchmem -benchtime=3x
node --experimental-strip-types web/scripts/benchmark-graph.mjs

See benchmark/README.md for isolated load tests. Short synthetic stages are not production capacity; requests are not unique visitors.

Original engine source: Blue Oak Model License 1.0.0. Third-party assets retain their own terms.