Skip to content

README.md

Latest commit

 

History

History
101 lines (85 loc) · 5.2 KB

README.md

File metadata and controls

101 lines (85 loc) · 5.2 KB
Raw

HiveBuzz performance benchmark

This benchmark records deterministic asset sizes, catalog bootstrap transfer, and a local Worker render proxy. Run it after a production build:

npm run build
npm run benchmark -- --label=current --output=benchmarks/current.json
npm run performance:check

The checked snapshots were collected on 2026-08-11 with Node 24.18.1. The runner warms the built Worker five times and records 30 home renders. Local render timings exclude the network and D1, so they are regression signals, not Cloudflare latency claims.

Result

Metric Before After Change
Home major initial byte budget, desktop 1,060,357 B 564,506 B 46.8% less
Home major initial byte budget, mobile 1,060,357 B 232,770 B 78.0% less
Home hero, desktop 767,648 B 469,382 B 38.9% less
Home hero, mobile 767,648 B 137,646 B 82.1% less
Static catalog bootstrap byte estimate 27,854 B, 2 call sites 460 B, 1 shared call site 98.3% less
Preloaded webfonts 146,464 B, 11 files 0 B, 0 files removed
Home HTML 58,998 B 51,692 B 12.4% less
Home HTML, gzip 9,935 B 8,752 B 11.9% less
Local Worker render p50 13.783 ms 4.625 ms 66.4% lower
Local Worker render p95 32.978 ms 11.365 ms 65.5% lower
Initial JavaScript, gzip 99,260 B 76,664 B 22.8% less

The major initial byte budget combines gzipped HTML, initial JavaScript and CSS, eagerly preloaded font bytes, the selected hero image, and the catalog bootstrap payload. It is a stable comparison budget rather than a browser HAR. The benchmark executes the API response path with a deterministic D1 stub to measure each raw JSON response, then multiplies that size by statically detected client call sites. A unit test separately proves concurrent callers share one in-flight request. HTTP compression, real D1 latency, Worker startup, and browser scheduling require production sampling.

npm run performance:check applies deterministic CI budgets to HTML, initial JavaScript and CSS, font preloads, catalog call sites, responsive home images, and the major initial byte estimate. Timing stays report only because shared CI latency is too noisy for a stable gate.

Three consecutive final after runs produced p50 values from 3.745 ms to 4.625 ms and p95 values from 5.692 ms to 11.365 ms, all below the before snapshot. This remains a local proxy rather than a production latency claim. The byte, request, cache, and query-shape results are deterministic and are the measured improvements used for the release gate.

Bottlenecks found

  1. The home page embedded the complete catalog in HTML, then HiveApp fetched the same 13,927 byte catalog again. SiteFooter made a second independent request, and every subpage fetched the full catalog solely to show service status.
  2. /api/releases read manifest JSON from D1, parsed it, validated it, and rebuilt every release on each uncached request. A fresh Worker isolate also performed catalog version synchronization before the read.
  3. The three 1774 by 887 hero backgrounds were served unchanged on mobile.
  4. Eleven Geist and Geist Mono font subsets were preloaded on every route.
  5. Request dependent metadata called headers() for a host that is fixed in production, adding work to every HTML render.

Changes

  • The initial catalog remains server rendered. A shared client loader now asks only for aggregate download counts and merges them into the trusted seed.
  • The count response reads only the downloads table and uses a short public cache window. The compatibility form of /api/releases remains available.
  • API responses bypass HTML nonce generation while retaining common security headers. HTML keeps its nonce based CSP unchanged.
  • Each hero now has a recompressed desktop asset and a 960 pixel mobile asset. Responsive high priority preloads expose the LCP image before CSS discovery.
  • The UI uses the system sans and monospace stacks, removing all webfont requests. The static production metadata no longer reads request headers.
  • The validated vinext 0.0.45 build removes the vulnerable image-size build dependency and emits a smaller initial client bundle while preserving every rendered and API regression test.
  • The footer reports catalog availability rather than making a broader system health claim from a briefly cached count response.
  • Regression tests enforce hero budgets, one count endpoint call site, no manifest database reads in that endpoint, no request dependent metadata, and no font preload or local filesystem URL in rendered HTML.

Production baseline

Before these changes, production sampling from Seoul observed:

Route Size Observed timing
/api/releases 13,927 B warm p50 about 430 ms, warm p95 about 750 ms, first observed response 3.995 s
/ 59,234 B warmed samples 255 to 372 ms TTFB; first samples 2.39 to 2.65 s
/guide 50,533 B warmed samples 141 to 487 ms TTFB
/contribute 26,176 B warmed samples 100 to 203 ms TTFB

Network timing is inherently noisy. A production after sample must use the deployed optimized version and the same route protocol; it is intentionally not filled from a local result.