# Large Dataset Plotting Benchmarks
This page is generated from the committed large-dataset plotting benchmark reference run.
## Methodology
- Output target: in-memory PNG byte generation only
- Dataset generation is excluded from all measured timings
- File I/O is excluded from all measured timings
- Boundaries:
- `render_only`: reuse a built plot object and measure an uncached render/export call without public-API reconstruction or prepared-frame image reuse
- `public_api_render`: reuse the input data, rebuild through the normal public API, then render/export
- Plot matrix: `line`, `scatter`, `histogram`, `heatmap`
- Python comparison target: `matplotlib` with the `Agg` backend
- Python `ruviz` benchmark runs use a release-built `maturin develop --release` extension
- Rust comparison target: `plotters` on the `public_api_render` boundary only
- `ruviz` PNG exports now use automatic raster fast paths for eligible cases:
- large monotonic solid lines without markers/error bars are reduced to a per-column envelope before stroking
- static histograms reuse prepared bins instead of re-binning on every render-only export
- nearest, non-annotated heatmaps rasterize directly to the output surface before PNG encoding
- Python host-side `ruviz` rendering now uses a persistent native plot handle and prepared plot instead of rebuilding a Rust plot from JSON on every render
- Python `render_only` timings bypass the prepared-frame image cache so they measure rasterization rather than cached PNG encoding
- Notebook widgets still ship JSON-friendly snapshots to the browser, but that snapshot path is no longer the default host render/export path
- Python full-mode runs cap very slow cases to a 60s per-case budget; the recorded warmup/measured counts reflect the effective counts used
- Rust `plotters` histogram timings reuse pre-binned bars, and the `plotters` heatmap path rasterizes the shared matrix to the fixed output canvas before PNG encoding
- wasm target: Chromium-only browser benchmark via Playwright
- Full-run warmup / measured iterations: `5` / `20`
## Why It Got Faster
The main change in this benchmark update is not a different benchmark harness. It is a different raster renderer path for large PNG exports.
What changed in `ruviz`:
- Large monotonic solid line series are reduced to a per-pixel-column envelope before stroking, so a `1M` point line on a `640px` canvas no longer pays to stroke every original segment.
- Static histograms now cache computed `HistogramData`, so `render_only` exports reuse prepared bins instead of re-running histogram binning on every frame.
- Nearest-neighbor, non-annotated heatmaps now render the final output surface directly and blit that image, instead of drawing one anti-aliased rectangle per source cell.
- The parallel line backend now emits a single polyline draw instead of thousands of two-point draw calls.
- Python host rendering now keeps a native Rust plot/prepared-plot handle alive across calls, so `render_png()`, `render_svg()`, `save()`, and `show()` no longer pay a Python JSON serialization + Rust JSON parse + plot reconstruction round-trip on every call.
Why those changes matter:
- The old line path scaled with source vertex count even when many samples collapsed onto the same output column.
- The old histogram path repeated statistical preprocessing inside hot render loops.
- The old heatmap path scaled with source cell count rather than output pixel count for raster exports.
- The old Python binding path spent a large share of its time turning Python state into JSON and then rebuilding a fresh Rust `Plot` before rendering anything.
The result is that current Rust PNG export timings mostly reflect output-resolution work for eligible raster cases, and current Python host-side timings reflect the renderer instead of the snapshot bridge much more closely than before.
## Scenario Matrix
| line | line_wave | 100k, 500k, 1m | 640x480 @ 100 DPI |
| scatter | scatter_cloud | 100k, 250k, 500k | 640x480 @ 100 DPI |
| histogram | histogram_signal | 100k, 1m, 5m | 640x480 @ 100 DPI |
| heatmap | heatmap_field | 512x512, 1024x1024, 2048x2048 | 640x640 @ 100 DPI |
## Environment
- Captured at: `2026-04-03T22:29:02Z`
- Git commit: `ada0957cdceb3e71bdddf080492cd17e5881c98a`
- Git branch: `bench/large-dataset-plotting`
- Host OS: `macOS-26.2-arm64-arm-64bit-Mach-O`
- Host machine: `arm64`
- Host processor: `arm`
- CPU count: `10`
- Python: `3.14.2`
- Rust: `rustc 1.94.1 (e408947bf 2026-03-25)`
- Bun: `1.3.5`
- Chromium: `145.0.7632.6`
Raw artifacts:
- [environment.json](../../benchmarks/plotting/results/reference/environment.json)
- [results.csv](../../benchmarks/plotting/results/reference/results.csv)
- [python.json](../../benchmarks/plotting/results/reference/python.json)
- [rust.json](../../benchmarks/plotting/results/reference/rust.json)
- [wasm.json](../../benchmarks/plotting/results/reference/wasm.json)
## Python: ruviz vs matplotlib (`render_only`)
| heatmap | 1024x1024 | 5.87 | 29.48 | 5.02x |
| heatmap | 2048x2048 | 7.32 | 61.83 | 8.44x |
| heatmap | 512x512 | 5.21 | 23.29 | 4.47x |
| histogram | 100k | 1.45 | 6.94 | 4.79x |
| histogram | 1m | 2.10 | 10.25 | 4.88x |
| histogram | 5m | 5.52 | 14.30 | 2.59x |
| line | 100k | 9.96 | 19.79 | 1.99x |
| line | 1m | 25.09 | 72.13 | 2.88x |
| line | 500k | 21.26 | 40.75 | 1.92x |
| scatter | 100k | 2.13 | 16.00 | 7.51x |
| scatter | 250k | 3.05 | 28.38 | 9.30x |
| scatter | 500k | 4.51 | 50.27 | 11.15x |
## Python: ruviz vs matplotlib (`public_api_render`)
| heatmap | 1024x1024 | 33.61 | 38.01 | 1.13x |
| heatmap | 2048x2048 | 122.75 | 72.79 | 0.59x |
| heatmap | 512x512 | 12.11 | 30.42 | 2.51x |
| histogram | 100k | 4.34 | 25.24 | 5.81x |
| histogram | 1m | 33.65 | 53.39 | 1.59x |
| histogram | 5m | 216.86 | 132.29 | 0.61x |
| line | 100k | 13.06 | 29.19 | 2.24x |
| line | 1m | 61.83 | 97.29 | 1.57x |
| line | 500k | 37.79 | 56.68 | 1.50x |
| scatter | 100k | 5.52 | 23.74 | 4.30x |
| scatter | 250k | 10.68 | 37.95 | 3.55x |
| scatter | 500k | 20.80 | 64.34 | 3.09x |
## ruviz cross-runtime medians (`render_only`)
| heatmap | 1024x1024 | 5.87 | 5.96 | 17.80 | b300137331a8 |
| heatmap | 2048x2048 | 7.32 | 7.90 | 18.70 | 2c5b8bacf8c3 |
| heatmap | 512x512 | 5.21 | 5.24 | 16.80 | d6e83388d86d |
| histogram | 100k | 1.45 | 1.51 | 5.50 | 52c1b6c47f7a |
| histogram | 1m | 2.10 | 2.28 | 6.10 | 4411c1d13a7c |
| histogram | 5m | 5.52 | 5.19 | 9.20 | 88dd6f1a74af |
| line | 100k | 9.96 | 13.58 | 19.70 | b011dead6d08 |
| line | 1m | 25.09 | 33.43 | 34.65 | a99f5c6c0498 |
| line | 500k | 21.26 | 27.01 | 27.60 | bb2c854f3c82 |
| scatter | 100k | 2.13 | 2.11 | 6.90 | a46d0038919e |
| scatter | 250k | 3.05 | 3.25 | 8.40 | 13bf7083712a |
| scatter | 500k | 4.51 | 4.52 | 10.60 | 6151ba3542b2 |
## ruviz cross-runtime medians (`public_api_render`)
| heatmap | 1024x1024 | 33.61 | 9.67 | 47.25 | b300137331a8 |
| heatmap | 2048x2048 | 122.75 | 20.92 | 225.70 | 2c5b8bacf8c3 |
| heatmap | 512x512 | 12.11 | 6.54 | 23.70 | d6e83388d86d |
| histogram | 100k | 4.34 | 3.04 | 8.55 | 52c1b6c47f7a |
| histogram | 1m | 33.65 | 19.40 | 40.90 | 4411c1d13a7c |
| histogram | 5m | 216.86 | 118.34 | 212.50 | 88dd6f1a74af |
| line | 100k | 13.06 | 14.01 | 22.90 | b011dead6d08 |
| line | 1m | 61.83 | 37.49 | 67.10 | a99f5c6c0498 |
| line | 500k | 37.79 | 27.78 | 45.70 | bb2c854f3c82 |
| scatter | 100k | 5.52 | 2.50 | 9.80 | a46d0038919e |
| scatter | 250k | 10.68 | 4.23 | 16.40 | 13bf7083712a |
| scatter | 500k | 20.80 | 6.65 | 27.00 | 6151ba3542b2 |
## Rust: ruviz vs plotters (`public_api_render`)
| heatmap | 1024x1024 | 9.67 | 2.36 | 0.24x |
| heatmap | 2048x2048 | 20.92 | 2.90 | 0.14x |
| heatmap | 512x512 | 6.54 | 2.51 | 0.38x |
| histogram | 100k | 3.04 | 0.73 | 0.24x |
| histogram | 1m | 19.40 | 0.72 | 0.04x |
| histogram | 5m | 118.34 | 0.73 | <0.01x |
| line | 100k | 14.01 | 1.71 | 0.12x |
| line | 1m | 37.49 | 11.19 | 0.30x |
| line | 500k | 27.78 | 5.93 | 0.21x |
| scatter | 100k | 2.50 | 2.63 | 1.05x |
| scatter | 250k | 4.23 | 5.72 | 1.35x |
| scatter | 500k | 6.65 | 10.71 | 1.61x |
## Rust render-only throughput
| heatmap | 1024x1024 | 176.01 M/s |
| heatmap | 2048x2048 | 530.90 M/s |
| heatmap | 512x512 | 50.02 M/s |
| histogram | 100k | 66.22 M/s |
| histogram | 1m | 439.34 M/s |
| histogram | 5m | 962.90 M/s |
| line | 100k | 7.36 M/s |
| line | 1m | 29.91 M/s |
| line | 500k | 18.52 M/s |
| scatter | 100k | 47.41 M/s |
| scatter | 250k | 76.82 M/s |
| scatter | 500k | 110.62 M/s |
## Notes
- These numbers are a reference snapshot from one machine and should be treated as comparative, not universal.
- Browser wasm timings include browser-side PNG generation, but not any disk writes or download flows.
- `render_only` is a reused-built-object benchmark, not a cached-frame benchmark: it avoids public-API reconstruction but still bypasses prepared-frame image reuse for `ruviz`.
- The remaining `plotters` gap on histogram and heatmap is partly semantic: `plotters` benchmarks pre-binned histogram bars and output-raster heatmap generation, while `ruviz` still includes its own plot-model setup and colorbar semantics on the public API path.