Each comparison supplies identical input to uPlot and another library. All reduction and grouping occur inside the libraries.
The left chart uses uPlot’s default linear-path decimation. It preserves pixel-column extrema, without averaging spikes or selecting only one point per bucket.
The right chart uses the labeled native algorithm. Opt-in settings are explicit. Sampling targets, activation thresholds, and grouping budgets keep their library defaults.
These examples are not speed tests. They show accuracy tradeoffs that timing benchmarks alone cannot establish.
Loading…
uPlot — default linear path, pixel-column extrema
Developer details — magnified paths
These troubleshooting views compare different rendering stages, not equivalent final images. Magnification exaggerates subpixel differences that can be invisible in the full-range charts.
uPlot — rendered path
Magnified uPlot path from the full-range chart, including its gap clip and pixel rounding.
Other library — processed data
Magnified polyline through the processed points, before renderer-specific rounding or further simplification. Unavailable when the library exposes no verified output.
How to read the comparison
All charts use a fixed −120…120 y range and the full input x range. Animations and point symbols are disabled.
X values represent minutes on a shared millisecond grid. Date-based libraries receive integer timestamps from January 1, 2024 (UTC).
Readouts distinguish input samples from processed output. They do not equate an unchanged input array with an accurate rendered path.
Average can attenuate or cancel spikes. Minimum and maximum retain only one direction. First and last can miss interior events.
LTTB can omit an extremum or change a pulse shape. Nonuniform spacing can expose weaknesses in count-based budgets.
Min/max alternatives and reduction-disabled modes are included. A comparison can preserve the feature; failure is not a requirement.
Results depend on plot width, DPR, bucket boundaries, and library version. No dataset-specific budget forces a failure.
Pixel-column extrema do not preserve every subpixel detail, timestamp, pulse shape, or signal statistic. No universal accuracy claim follows from these examples.
The benchmark uses separate accuracy-first settings. FusionTime’s one-minute mode is exact only for suitable minute-aligned, one-sample-per-bin data.