Skip to content

v3.3.0 — Major performance overhaul: 8x faster exports, real multi-core parallelism

Latest

Choose a tag to compare

@asinay asinay released this 29 Sep 03:26

SysPerfSight

This release is headlined by a major performance overhaul: large report exports that used to hang indefinitely (or take multiple minutes) now reliably finish in seconds, with real multi-core parallelism and a live progress indicator. The rest is a batch of reliability and UX fixes accumulated since v3.2.0.

Performance

  • Exports on large reports no longer hang indefinitely — parsing, analysis, and output generation now run through a thread pool with timeouts, so a pathological file fails cleanly with a 504 instead of blocking the server forever
  • Per-section analyzers now run in real separate worker processes instead of threads, giving genuine multi-core parallelism (analyzers are pure CPU work with nothing to await, so threads alone mostly serialize on the GIL) — cross-section synthesis also now runs concurrently with the analyzers instead of waiting for them
  • Fixed a severe iostat slowdown on systems with many disk devices (multipath LUNs, lots of local disks): a report with 75 devices was generating 528 individual Plotly chart-trace calls plus repeated full-dataframe scans — iostat: 111s → 6.6s, total export: 116s → 14.7s (~8x), output size: ~121MB → ~85MB on the same real-world file. Charts now cap to the top 15 devices by %util (readable, and the summary table/insights still cover every device)
  • mgstat, iostat, sar-d, and %SS gained the same row/interval/snapshot downsampling cap already used by vmstat/perfmon, bounding chart size on long captures
  • Export progress is now visible: a live "Generating report… Ns elapsed (N sections, X MB)" status shows under the export buttons instead of just a static "Generating…" button label
  • Per-stage timing is logged server-side (docker logs) for future diagnosis without needing the report file itself — see PERFORMANCE.md

Fixes

  • Uploads no longer silently fail on files re-saved by a browser ("Save Page As" instead of the raw source) — this reserializes the section-heading markup into a structurally different but visually identical form; it's now detected and normalized automatically
  • A 0-section parse now returns an actionable 400 error (with a diagnostic hint) instead of a false "success" with an empty section list
  • Error messages are now readable even when a reverse proxy rejects a request before it reaches the app (e.g. a 413 from an upload size limit) instead of an opaque Unexpected token '<' from trying to parse an HTML error page as JSON
  • Fixed the Docker image still referencing the pre-rename pbuttons_parser.py
  • Renamed the published GHCR image from sysperf-sight to sysperfsight to match the repo/container name

UI

  • Added a FAQ page and made all internal fetch/link URLs relative, so the app keeps working when reverse-proxied under a subpath
  • The generated report's header now matches the app's own header style (wordmark, colors, dark-mode toggle)
  • General UI redesign as an ops tool, with leftover pButtons-era branding cleaned up