docs(benchmarks): the benchmark page highlights specsolve, projects the sizes a budget refused, and compares every library with it in one table

[FBumann]

Prompt: I would rather like you to propose a better plotting for our benchmarks.

Prompt: Absolute numbers are important! But maybe we can add a toggle for log or linear scaling? … We should highlight specsolve! — A as the lead + E is what i want. — I would like to have the axes both scaled properly — Could we add ticked lines as extrapolations? — Use a blue to green color for the heatmap — We could use our own brands violet…? — Should we unselect the pure matrix ones? They are merely an oracle and probably not comparable…? — Yes, style them neutral. And probably also in the heatmap…?

Note

The following content was generated by AI.

The benchmark page leads with absolute cost against size. specsolve is a thick line in the site's brand violet; the other libraries are thinner, and the two hand-written matrix arms are one neutral "matrix floor". One switch flips both axes between log–log and linear–linear, and dashed lines project the sizes a budget refused. A single ratio table, shaded blue to green, replaces the per-panel tables.

What the page draws

• Panels. One panel per model and solver: each library's cost against size, with the median line and the Q1–Q3 band. specsolve's line is 3.5 px and the libraries' lines are 2 px. specsolve is drawn last and named above its last measured point. Every measured point is a dot ringed in the surface colour.
• Matrix floor. gurobipy-matrix and highspy-matrix are a hand-written CSR handed to the solver's bulk API, the limit a modelling library can approach. They are drawn as one thin grey line with no band, behind one legend entry, and they never share a panel. In the table they are one unshaded "matrix floor" column at the right, because being faster than specsolve is what a floor is for. A note on the page says what the floor is. gurobipy-loop stays a library.
• Axes. One switch for log–log or linear–linear. Mixed axes are left out: linear time over log size draws linear growth as an upward curve.
• Projections. A dashed line goes from a library's last measurement to each size its time or memory budget refused, at the log–log slope of the last measured step, floored at 1. A hollow ring marks each projected size. The axes scale to the measurements, so a steep projection is clipped at the top. The tooltip shows ≈ value (projected).
• Table. specsolve's own cost, then each library's cost divided by specsolve's at the same size: green where specsolve is faster, blue where the other library is, grey within 10%. A cell's hover gives the library's own number. A refused cell keeps its >30 s or >16 GB label and gives the projection on hover.
• Spec. The JSON spec gains highlight, floor (its label and members), and the ink and paint colour lists. The ratios, the projections, the floor and the highlighting all read from it.

Colour

• Palette. specsolve brand violet (#6d28d9; in dark mode #9670e8, because the site's #a78bfa is too light for a chart line), linopy gold, pyomo blue, gurobipy-loop red (Gurobi's colour), and the matrix floor neutral grey (#6b6a61; in dark mode #a3a196).
• Validation. The four library colours pass the dataviz skill's validator in both themes, with worst adjacent CVD ΔE 26.0 in light mode and 24.7 in dark (protan). The libraries other than specsolve are each hue mixed into the surface, 70% in light mode and 85% in dark.
• Brand colours. linopy, Pyomo and HiGHS have no signature colour I could verify.
• Table. The blue–green pair passes deutan (ΔE ≈ 20) but is weak for tritanopia (ΔE ≈ 5). Every cell prints its ratio, which carries the reading there.

What was verified, and what was not

• Browser: headless Chromium in light and dark, with the jsdelivr files served from the npm tarball as in docs(benchmarks): the benchmark charts read a published JSON table of measurements and label our engine specsolve #1851. 5 panels and 20 table rows. Switching to linear–linear, the width ladder and peak RSS, and hiding pyomo, gives 1 panel, 4 rows, and each refused cell's projection on hover. No console errors.
• Gates: pixi run docs-build (strict) passes and ships the page and its JSON. The diff touches no Python. CI was green on the previous head.
• Not checked: loading from the real jsdelivr. The Read the Docs preview of this PR is the first real load.

Stack and what was not done

• Base: stacked on docs(benchmarks): the benchmark charts read a published JSON table of measurements and label our engine specsolve #1851, which carries the JSON data (the numbers re-measured on 0.6.0) and the spec this page reads.
• Not done: no change to bench/plot.py or the data model. The per-panel absolute tables are gone; the specsolve column carries absolute values, and every other value is in the cell hovers and chart tooltips. The panels stay three to a row.

🤖 Generated with Claude Code

https://claude.ai/code/session_01K42KwZJtqPPhf5hDzUD91o

[read-the-docs-community]

Documentation build overview

📚 specsolve | 🛠️ Build #34979039 | 📁 Comparing 1fb99c2 against latest (79258b1)

  🔍 Preview build  

2 files changed

± about/benchmarks-scaling.html
± about/changelog/index.html

[codspeed]

Merging this PR will not alter performance

✅ 67 untouched benchmarks
⏩ 307 skipped benchmarks¹

---

_{Comparing claude/benchmark-chart-design (1fb99c2) with claude/exciting-ritchie-ap020w (49cf2be)}

Footnotes

1. 307 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩