refract

refract

A grammar-driven plotting library for Go: one model, many backends, runs everywhere — built on the GoGPU stack.

Status: v1.0.0, released. Every milestone through v1.0 has shipped, and the v1 API audit is in: what it asked to change before the freeze has changed. The API is frozen at the v1.0.0 tag and follows semver from here, so a breaking change means a major version and a deprecation cycle precedes it. The nested modules are tagged alongside: backend/gg and backend/window at v1.0.0, the opt-in GPU tier at v0.1.0, and the Arrow adapter at arrow/v18.0.0, whose major is Arrow's. See CONCEPT.md for the design and the road ahead.

The name is the thesis: one beam enters a prism, a spectrum comes out. One chart specification enters refract, a spectrum of output formats comes out.

---

What it is

One declarative chart specification, rendered through interchangeable backends. The core is pure Go with no dependencies at all — not "no cgo", literally nothing outside the standard library — and emits both vector formats, SVG and PDF. Add one module and the same specification renders to PNG and JPEG through gogpu/gg, still with CGO_ENABLED=0.

	Dependencies	Output
github.com/timzifer/refract	stdlib only	SVG, PDF, browser canvas
github.com/timzifer/refract/backend/gg	GoGPU (gg), x/image — zero CGO	PNG, JPEG, an in-memory surface
github.com/timzifer/refract/backend/window	GoGPU (gogpu, gg) — zero CGO	a native window
github.com/timzifer/refract/backend/gg/gpu	GoGPU (gg/gpu, wgpu) — zero CGO	— (switches the GPU tier on)
github.com/timzifer/refract/arrow/v18	apache/arrow-go — zero CGO	— (a data source)

The browser is in the core too, because it needs nothing to be: a canvas 2D context is reached through syscall/js, which is the standard library (ADR 0017). Everything else is behind the same ir.Backend interface, in a module of its own, so what a program links is what it asked for: a server that renders SVG links nothing but the standard library, and a desktop program that opens a window links a window layer.

Install

go get github.com/timzifer/refract                  # core: SVG and PDF, stdlib only
go get github.com/timzifer/refract/backend/gg       # raster: PNG and JPEG
go get github.com/timzifer/refract/backend/window   # a native window
go get github.com/timzifer/refract/backend/gg/gpu   # optional: the GPU tier
go get github.com/timzifer/refract/arrow/v18        # optional: plot Arrow data

Go 1.25 or newer (why).

Quick start

package main

import (
	"log"
	"math"

	"github.com/timzifer/refract"
	"github.com/timzifer/refract/geom"
	"github.com/timzifer/refract/palette"
	"github.com/timzifer/refract/scale"
	"github.com/timzifer/refract/theme"
)

func main() {
	xs := make([]float64, 200)
	ys := make([]float64, 200)
	for i := range xs {
		xs[i] = float64(i) / 20
		ys[i] = math.Sin(xs[i])
	}

	p := refract.New(
		refract.Theme(theme.Dark),
		refract.Size(800, 400),
		refract.Title("Signal"),
		refract.YTitle("amplitude"),
	)
	p.X(scale.Linear(scale.Nice()))
	p.Y(scale.Linear(scale.Nice()))
	p.Add(geom.Line(
		refract.Float64Columns(map[string][]float64{"x": xs, "y": ys}),
		geom.X("x"), geom.Y("y"),
		geom.Color(palette.Blue),
		geom.Tension(0.4),
	))

	if err := p.Render(refract.SVG("signal.svg")); err != nil {
		log.Fatal(err)
	}
}

For PDF or raster, swap the target — nothing else changes:

err := p.Render(refract.PDF("signal.pdf"))          // still stdlib only

import ggbackend "github.com/timzifer/refract/backend/gg"
err := p.Render(ggbackend.PNG("signal.png"))

A runnable version is in examples/signal.

Gallery

Every figure below is rendered by backend/gg/cmd/gallery and re-checked in CI, so a picture here cannot drift away from the code that produced it.

What it does

• Scales — linear, time, log, symlog and ordinal/categorical. Linear tick placement uses extended Wilkinson (Talbot, Lin & Hanrahan 2010), so axis labels come out round rather than merely evenly spaced. Time ticks step in calendar units. Log and symlog subdivide each decade with unlabelled minor ticks; symlog is linear near zero, so signed data spanning orders of magnitude is plottable at all.
• Geoms — Line (optionally tension-smoothed), Scatter (six marker shapes), Bar, Area (to a baseline, or a band between two series), Step (pre/mid/post), Boxplot (Tukey whiskers, type-7 quartiles, outliers), and Rect — one box per row, bounded by the row rather than by a baseline, which is what a heatmap, a gantt bar, a candle and a waterfall step all are.
• Distribution marks — Histogram, Violin, Ridgeline, Hexbin, Beeswarm, ECDF and Trend. Each is a pure function in stat/ — a 1-D binner, a Gaussian KDE with Silverman's bandwidth rule, a hexagonal lattice, an empirical CDF, locally weighted regression — with a determinism test, drawn by a mark that trains its axis on the summary rather than on the rows (ADR 0028).
• Series in one layer — geom.GroupBy splits a long table into N series drawn by one layer, each with its own colour and its own legend entry.
• Position adjustments — geom.Stack (from zero, to 100 %, about a silhouette, or with the streamgraph's wiggle) and geom.Dodge (side by side). The offsets are derived while the scales are trained, so a stacked axis reaches the total rather than the tallest single value (ADR 0019).
• Colour — a qualitative palette per chart, plus colour scales bound to a column by geom.ColorBy: a sequential or diverging ramp for a quantity, or scale.Qualitative for categories. Which guide the layer contributes follows from which it was handed — a ramp gets a colourbar, a palette gets one legend entry per category (ADR 0020). Ramps interpolate in linear light, so a gradient has no dark band through its middle.
• Coordinate systems — coord.Cartesian is the identity and the default; coord.Polar wraps one axis around a circle and reads the other as a radius, which turns the marks that already exist into pie, donut, radar, rose, wind rose and gauge. A scale still maps a value into an interval — the coord decides what the interval means — so no geom and no scale changed shape for it, and the arcs are cubics because the IR has always had those (ADR 0018). A slice's inner and outer radius are columns like its share is (geom.X and geom.X2), and geom.ExplodeBy breaks one out of the ring without changing what it says (ADR 0026).
• Size — geom.SizeBy reads a column through scale.Size: the bubble chart. The mapping is by area, not radius, so doubling a value multiplies the diameter by √2 and two bubbles compare the way a reader already reads them. The layer contributes a third guide kind — a ladder of sample marks — beside the legend and the colourbar (ADR 0027).
• Missing data — one explicit policy per layer (gap, interpolate, error), covering both NaN/Inf and values a scale has no position for, such as zero on a log axis.
• Annotations — HLine, VLine, HBand, VBand, Segment, Region and Note. They take values rather than a data source, because there is no column behind "the SLO is 200ms", and they extend the axis so the threshold is in view even when the data is nowhere near it.
• Small multiples and subplots — facet.Wrap and facet.Grid split one plot by a column; refract.NewGrid puts different plots on one canvas. Both go through one constraint solver, so panels are the same size and their axes line up (ADR 0010).
• Chart furniture — axes, grid, tick labels with collision avoidance, chart and axis titles, and one guide column carrying a legend, colourbars and size keys, stacked in that order and measured by one solver.
• Themes — light and dark, built from a dozen tokens rather than fifty fields, with a colourblind-safe (Okabe-Ito) default palette and perceptually uniform sequential ramps (Viridis, Cividis, Magma). Theme.With edits one; theme.Register and theme.ByName resolve one from a config file.
• Big data — a layer with more rows than the plot has pixels reduces itself before it draws: stat.LTTB for a line, min/max per pixel column for a staircase or a band, density binning to an image for a point cloud, or a geom.Hexbin when the counts themselves are the answer. It happens when the chart is drawn, never when the scales are trained, so the axes still report the data rather than the subset that survived (ADR 0011).
• Parallel panels — a facet or a grid builds its panels on separate goroutines and replays them in panel order, so the output is byte-identical to a serial render (ADR 0012).
• Data — columnar and batch-oriented, carrying numeric, time and categorical columns. A []float64-backed source is borrowed, never copied, and so is a null-free float64 column read straight out of an Apache Arrow record through the optional refract/arrow/v18 module.
• Interaction — Plot.On registers handlers for hover, click, zoom and pan; Plot.Live draws into a surface that can be redrawn; refract.Input is the state machine that turns raw pointer input into those, and Live.Bind drives it from a DOM element. Hit-testing runs over the marks a render emitted rather than over a second copy of every geom's projection, and Live.TrackRows makes a hit name the source row behind the mark (ADR 0015).
• Live data — data.Stream is appended to from any goroutine and frozen between frames, and a redraw repaints only what changed (ADR 0016).
• A chart as JSON — a *Plot marshals to a Vega-Lite-shaped document and reads back as the same chart (ADR 0014).
• Accessibility — a chart's title becomes an SVG <title> with role="img", a PDF document title and a canvas aria-label; Plot.Describe writes the <desc> a screen reader announces after it; Plot.DataTable writes the rows as an HTML table; and theme.Redundant tells layers apart by dash and shape as well as by colour (ADR 0024).
• Notation in labels — optional and pluggable, with a TeX subset built in. A label is measured as it will be drawn, in every place a chart writes one (ADR 0023).
• Responsive charts — refract.Responsive scales a theme with the size the chart is drawn at, and Live.Resize is how a surface says its size changed (ADR 0025).
• Backends — three built-in emitters — SVG, PDF and a browser canvas — the gg raster adapter, a native window, and an opt-in GPU tier.

Deliberately not here: geographic projections, the relational layouts — sankey, chord, treemap, sunburst — contour and QQ plots, animation, and 3D. They are past v1.0 in CONCEPT.md §14, and docs/chart-types.md says what each one would need.

Categories, distributions and orders of magnitude

src := refract.NewTable().
    String("region", []string{"north", "south", "east", "west"}).
    Float64("sales", []float64{18, 42, 31, 25})

p := refract.New(refract.Size(700, 400), refract.Title("Sales by region"))
p.X(scale.Ordinal())                        // equal slots, one per category
p.Y(scale.Linear(scale.Nice(), scale.Zero()))
p.Add(geom.Bar(src,
    geom.X("region"), geom.Y("sales"),
    geom.ColorBy("sales", scale.Sequential(palette.Viridis)),
))

An ordinal axis is a band scale: it tells the bar how wide to be, rather than the bar guessing from the spacing of the data. The same applies to geom.Boxplot. For data that spans decades, swap in scale.Log(scale.LogNice()) — or scale.SymLog() when it also crosses zero.

A runnable version, together with a boxplot over the same kind of data, is in examples/categories.

Series in one layer, stacked or side by side

A table with a series column is one layer, not N. geom.GroupBy splits it, and the position adjustments are defined over the groups it makes:

// quarter, product, revenue — twelve rows, one per (quarter, product) pair
p.X(scale.Ordinal())
p.Y(scale.Linear(scale.Nice(), scale.Zero()))
p.Add(geom.Bar(src,
    geom.X("quarter"), geom.Y("revenue"),
    geom.GroupBy("product"),
    geom.ColorBy("product", scale.Qualitative(palette.OkabeIto)),
))

A grouped bar stacks from the baseline up, because that is what a bar chart with a series column means. geom.Dodge(0.1) puts the products side by side instead, geom.Stack(geom.StackFill) makes it a 100 % chart, and geom.Stack(geom.StackWiggle) over geom.Area is a streamgraph. The axis is trained on what will be drawn rather than on the column, so a stacked axis reaches the total; each segment is its own shape, so a pointer lands on the segment and Live.TrackRows names the row behind it.

The legend names every series. One swatch per layer could not, which is why a layer contributes as many entries as it has to (ADR 0020).

A pie is a stacked bar in a different coordinate system

A scale maps a value into an interval; a coord decides what that interval means. coord.Cartesian — the default — says it is a distance along an edge of the plot. coord.Polar says one of the two intervals is an angle and the other a radius, and the marks that were already there draw the family that was missing:

p := refract.New(
    refract.Coord(coord.Polar(coord.Theta(coord.FromY), coord.Hole(0.45))),
    refract.Theme(theme.Light.With(
        theme.Grid(false, false), theme.AxisLines(false, false), theme.Ticks(false, false))),
)
p.X(scale.Linear())          // one slot, filling the radius
p.Y(scale.Linear())          // the stacked total, filling the circle
p.Add(geom.Bar(src, geom.X("all"), geom.Y("share"),
    geom.GroupBy("browser"),
    geom.ColorBy("browser", scale.Qualitative(palette.OkabeIto))))

That is the whole of the donut above: the same geom.Bar layer that draws a stacked bar chart in Cartesian, with θ taken from the Y axis instead. The ring closes into a full circle because a stacked domain ends at the total — which is why neither scale is niced — and the hole is where the radial scale starts, so it is an annulus rather than a circle of background painted over the middle.

A slice's radii are dimensions, and a slice can leave the ring

The donut above carries one number per slice: its share, which is the angle. Two more are already there for the taking, because the radial axis is an axis like any other. geom.X and geom.X2 name a mark's two edges on it — the pair a gantt bar has used since v0.7 — so a slice starts and stops where its row says:

p := refract.New(refract.Coord(coord.Pie(coord.Radius(0.95))), refract.Theme(bare))
p.X(scale.Linear(scale.Domain(0, 1)))   // the radius: 1 is the whole budget
p.Y(scale.Linear())                     // the angle: the stacked share
p.Add(geom.Bar(src,
    geom.X("floor"), geom.X2("used"),   // where the slice starts and stops
    geom.Y("share"),                    // how far round it goes
    geom.GroupBy("team"),
    geom.ExplodeBy("pull"),             // and which of them leaves the ring
    geom.ColorBy("team", scale.Qualitative(palette.OkabeIto))))

That is the figure above: three numbers per slice, one layer, no new mark. coord.Pie() and coord.Donut(f) are sugar for the polar recipe and describe themselves as the polar coord they are.

geom.Explode(f) breaks every mark of a layer out of the middle by a fraction of the outer radius; geom.ExplodeBy(col) reads that fraction per row, which is what pulls one slice out and leaves the rest where they were. It is a displacement rather than a longer radius, and that is the whole point: the slice still says what it said, the gap shows where it came from, and a pointer follows it out — the path a geom hands the backend is the path that gets indexed, so a hit in the gap finds nothing. A coord with no middle to move away from — coord.Cartesian — ignores it rather than inventing a direction, which is why every golden file in the repository is unchanged by an option every geom now accepts (ADR 0026).

A radar is geom.Line or geom.Area over an ordinal angular axis, with coord.Chord() for sides that are straight and geom.Closed(true) for a contour that comes back to the first axis. A rose, a wind rose and a gauge are bars; a polar boxplot is a boxplot. None of them is a new geom, which is the point of having a stage rather than a shape (ADR 0018).

The stage costs an existing chart nothing: Cartesian is the identity, and every golden file and every figure in this README is unchanged by it. What it does change is what a pointer can be told — a hit is inverted back through the coord before the scales see it, so a pointer over a slice reports the value the slice stands for rather than a pixel, and Live.TrackRows names the row. Concentric rings replace horizontal grid lines and the tick labels go round the outside; the coord reports that geometry and render still strokes it, because render is the only package that knows the drawing order of a chart.

Decimation is deliberately off under a polar coord. stat.LTTB buckets by pixel column and a bucket of equal angle is not a bucket of equal width, so the coord reports that it does not decimate rather than have a reduction measure something it was not designed for. Nothing polar is a big-data chart, so this costs nothing real.

A runnable version of the donut, the radar, a wind rose, a gauge and the broken-out donut above is in examples/polar.

Boxes bounded by their own row

geom.Rect occupies an arbitrary [x0,x1] × [y0,y1] per row — the mark a bar is not, because a bar always touches the baseline. An edge no column names is the slot the axis implies, so a heatmap is a rect and a ramp:

p.X(scale.Ordinal(scale.OrdinalPadding(0)))
p.Y(scale.Ordinal(scale.OrdinalPadding(0)))
p.Add(geom.Rect(src, geom.X("day"), geom.Y("hour"),
    geom.ColorBy("calls", scale.Sequential(palette.Viridis))))

and a gantt bar, which knows where it starts and stops, names both:

p.X(scale.Time())
p.Y(scale.Ordinal())
p.Add(geom.Rect(src, geom.X("from"), geom.X2("to"), geom.Y("task")))

Candlestick, waterfall, waffle and calendar are the same mark with different columns — see docs/chart-types.md.

A runnable version of all four charts is in examples/groups.

Distributions

Seven marks that summarise a column rather than plotting it. Each is a pure function in stat/ with a determinism test, and each trains its axis on the summary — a histogram's Y axis holds counts that appear nowhere in the table, an ECDF's holds a fraction it computed (ADR 0028).

p.Add(geom.Histogram(src, geom.X("latency")))              // bins chosen by Freedman–Diaconis
p.Add(geom.Histogram(src, geom.X("latency"), geom.Bins(40), geom.BinRange(0, 500)))

A violin draws the shape a boxplot summarises away, one per slot and — given a series column — one per series within it:

p.X(scale.Ordinal())
p.Add(geom.Violin(src, geom.X("service"), geom.Y("latency"), geom.GroupBy("region")))

A ridgeline is the same estimate laid out down a categorical axis, overlapping on purpose: twenty little density panels are twenty comparisons a reader has to carry between them, and twenty ridges are one picture.

p.Y(scale.Ordinal())
p.Add(geom.Ridgeline(src, geom.X("temperature"), geom.Y("month"), geom.Overlap(2)))

A swarm shows every observation and hides none of them, deterministically — no jitter, so the same data draws the same picture on every machine and every frame. An ECDF shows a distribution with no parameter in it at all, and takes a series column so several can be compared without overplotting:

p.Add(geom.Beeswarm(src, geom.X("cohort"), geom.Y("score")))
p.Add(geom.ECDF(src, geom.X("score"), geom.GroupBy("cohort")))

A hexbin is the third answer to overplotting, beside decimation and the density raster: a hexagon has six neighbours all the same distance away, so a cloud binned into one grows none of the crosses and seams a square grid does.

p.Add(geom.Hexbin(src, geom.X("x"), geom.Y("y"), geom.DensityCells(8)))

And a trend line goes on top of a scatter — locally weighted by default, so it follows the data rather than assuming a shape:

p.Add(
    geom.Scatter(src, geom.X("x"), geom.Y("y")),
    geom.Trend(src, geom.X("x"), geom.Y("y"), geom.Span(0.4)),
    geom.Trend(src, geom.X("x"), geom.Y("y"), geom.Smooth(geom.LinearFit)),
)

All seven, over samples that make the point, are in examples/distributions.

Bubbles: a third channel

geom.SizeBy gives every mark its size from a column. The scale maps by area rather than by radius, because a reader compares two circles by how much ink is in them — so a value twice another's is drawn with twice the ink and √2 times the diameter, and the layer contributes a key of sample marks beside the legend and the colourbar:

p.Add(geom.Scatter(src,
    geom.X("gdp_per_capita"), geom.Y("life_expectancy"),
    geom.SizeBy("population", scale.Size()),
    geom.ColorBy("continent", scale.Qualitative(palette.OkabeIto)),
))

A sized layer draws circles rather than markers, and that is the IR's doing rather than a preference: ir.Backend.Markers carries one style per drawing call, so a per-row size would be a call per row. One path per colour with a circle per subpath is one call per colour — and it gives a pointer the bubble it is actually inside rather than the nearest centre (ADR 0027).

The chart is in examples/distributions too.

Small multiples

p := refract.New(refract.Size(900, 520), refract.Title("Throughput by region"))
p.Add(
    geom.Line(src, geom.X("hour"), geom.Y("rps"), geom.Label("throughput")),
    geom.HLine(60, geom.Label("target")),          // no data: drawn on every panel
)
p.Facet(facet.Wrap("region", facet.Columns(3)))

Panels share their scales by default, which is what makes small multiples comparable at a glance. facet.FreeX, facet.FreeY and facet.Free give each panel its own — a deliberate choice, because a reader who does not notice the axes changed will read the panels as comparable when they are not.

For unrelated charts on one canvas, build a grid of plots instead:

g := refract.NewGrid(2, refract.GridSize(900, 560), refract.GridTitle("Fleet"))
g.Add(latency, throughput, errors, saturation)
err := g.Render(refract.PDF("overview.pdf"))

A runnable version of both, with annotations and PDF output, is in examples/dashboard.

A band at the edge, on the same axis

Some of what a chart shows is not on its other axis at all: a strip of machine states under a speed trace, a rug of event times, a ribbon of shifts, a key or a marginal distribution beside the panel. A track is a band at an edge of the plot area that shares the axis it runs along and carries a scale of its own across it — an ordinal one under a linear panel, which is the case it exists for.

p := refract.New(refract.Size(900, 480), refract.Title("Line 3"))
p.X(scale.Time()).Y(scale.Linear(scale.Zero()))
p.Add(geom.Line(speed, geom.X("t"), geom.Y("speed")))

p.Track(refract.Bottom, refract.TrackSize(48)).
    Add(geom.Rect(states, geom.X("start"), geom.X2("end"), geom.Y("state"),
        geom.ColorBy("state", scale.Qualitative(palette.Default))))

refract.Bottom and refract.Top are grid rows and share the plot's X; refract.Left and refract.Right are grid columns and share its Y. Bands on two edges at once are fine — the corner between them is simply empty.

The thickness comes out of the panel, not out of the panel's domain: the axis the track does not share is identical with the track and without it, so scale.Zero() still means what it says. The axis it does share is trained by both, because it is one axis — a rug of event times widens the time axis to cover the events, which is the reason to draw them against it.

The track and the panel hold the same scale object for that axis, so a zoom is one zoom rather than two that agree, and a pointer over a state bar reports its layer and its row like any other mark. Its lanes do not zoom, because half a category is not a view of anything.

For the same shape across separate plots, stack them in a one-column grid on one scale object — refract.GridRowHeights(0, 48) makes the second row a strip and refract.GridSharedX(true) writes the time axis once, under the bottom row, with GridColWidths and GridSharedY the same turned a quarter turn. That path renders; interaction is what a track is for.

A million rows

p := refract.New(refract.Size(800, 500), refract.Title("A million samples"))
p.Add(geom.Line(src, geom.X("i"), geom.Y("v")))   // nothing else needed

That renders in about 60 ms into under 30 kB of SVG. Drawing every row takes six times as long and produces 15 MB — of a picture that is 800 pixels wide, so the extra 999,000 vertices land on top of each other.

The layer sees how many rows it has against how wide the plot is and reduces itself accordingly: LTTB for a line, min/max per pixel column for a step or a band, a density raster for a scatter dense enough that its markers would bury one another. Override it per layer when the default is not what you want:

geom.Line(src, geom.X("i"), geom.Y("v"), geom.Decimate(geom.MinMax))    // keep every spike
geom.Line(src, geom.X("i"), geom.Y("v"), geom.Decimate(geom.NoDecimation)) // every row
geom.Scatter(src, geom.X("x"), geom.Y("y"), geom.Budget(4000))          // at most 4000 marks

The reduction happens when the chart is drawn, not when its scales are trained, so the axes are the data's either way — a spike survives the reduction and the axis still reaches it.

The same milestone made a redrawn chart cheap: everything sized by the data comes from a pool, so a steady-state frame over a million rows costs the same handful of allocations as one over a thousand. There is a test that fails if that stops being true.

A runnable version — two million samples with a spike and a dropout in them, and a million-point cloud — is in examples/bigdata.

Interactive, in a browser

The same model that renders SVG on a server draws on a <canvas>, with the pointer reporting what it is over:

//go:build js && wasm

p.On(refract.Hover, func(ev refract.Event) {
	if ev.Found {
		readout.Set("textContent", fmt.Sprintf("%s: %.2f, %.2f", ev.Series(), ev.Hit.X, ev.Hit.Y))
	}
})

live, err := p.Live(canvas.Element(el))   // a surface, redrawn
defer live.Close()
live.Draw()
defer live.Bind(el)()                     // pointer, wheel zoom, drag pan, double-click reset

Hit-testing works over the marks the render actually emitted, so it is right for every geom — including a decimated one, where the rows you can point at are exactly the rows on screen (ADR 0015). Zoom and pan are arithmetic on the scales, so the value under the pointer stays under the pointer on a log or a time axis as much as on a linear one.

Turn on row identity when a hit has to name a row rather than describe a point — highlighting the matching row of a table beside the chart is the case:

live.TrackRows(true)

p.On(refract.Hover, func(ev refract.Event) {
	if ev.Found && ev.Hit.Row >= 0 {
		highlightTableRow(ev.Hit.Row)   // a row of the table you handed in
	}
})

It is off by default and costs a position and a row number per mark; it does not cost per-frame allocations, and CI pins that. Decimation is not in the way — LTTB and min/max keep real rows — and neither is faceting, whose per-panel cuts are resolved back to the table you passed. A mark that no single row is behind — a boxplot's box, a density raster, an interpolated point across a gap — reports -1 rather than a plausible neighbour.

A runnable version is in examples/web:

GOOS=js GOARCH=wasm go build -o examples/web/chart.wasm ./examples/web
cp "$(go env GOROOT)/lib/wasm/wasm_exec.js" examples/web/

Interactive, in a window

The same model again, on the desktop:

import (
	"github.com/timzifer/refract/backend/window"
	"github.com/timzifer/refract/backend/window/show"
)

func main() {
	p := refract.New(refract.Responsive(true), refract.Title("Signal"))
	p.Add(geom.Line(src, geom.X("t"), geom.Y("y")))

	// Hover, drag to pan, wheel to zoom about the pointer, double click to reset.
	log.Fatal(show.Plot(p, window.Title("Signal"), window.Size(900, 560)))
}

The window comes from gogpu/gogpu — Windows, macOS, X11 and Wayland, no cgo — and the chart is drawn by the same CPU rasterizer that writes your PNGs, presented as one texture per changed frame. So a window shows exactly what a file would; there is one implementation of every mark rather than two that disagree (ADR 0021).

It is also cheap when nothing is happening: the loop blocks on the operating system's event queue, refract paints nothing when a frame is identical to the last, and the window uploads no texture when the pixels have not changed.

The steering — is this move a hover or a drag, was that release a click — is refract.Input, in the core, and it is the same state machine Live.Bind uses in a browser. Drive it yourself if you want different controls:

in := live.Input()
in.Down(x, y); in.Move(x, y); in.Up(x, y)   // press, pan, release
in.Wheel(x, y, deltaY)                      // zoom about the pointer
in.Resize(w, h)                             // lay out again at a new size

A runnable version is backend/window/cmd/demo:

cd backend/window && go run ./cmd/demo

The GPU tier, opt-in

import _ "github.com/timzifer/refract/backend/gg/gpu"

That import registers gg's GPU accelerator, and every chart rasterized afterwards — in a window or into a file — uses it. It is a module of its own so that the import is the opt-in: backend/gg never links wgpu, and a program that wants a PNG on a server links no GPU stack at all (ADR 0022). On a machine with no usable device the registration fails quietly and gg falls back to the CPU, so the chart still renders; gpu.Enabled() says which way it went.

It stays opt-in beta past v1.0. For server-side stills the CPU rasterizer and the vector emitters are the supported path.

Labels that are notation

p := refract.New(
	refract.Math(mathtext.TeX()),
	refract.YTitle(`flux $F_\nu$ ($\mathrm{W\,m^{-2}\,Hz^{-1}}$)`),
	refract.Title(`decay of $N_0e^{-\lambda t}$`),
)

$…$ is set as notation and everything around it as text. The subset is the one a chart label actually needs — scripts, \frac, \sqrt, \bar, \mathrm, the spacing commands, and a table of symbols — and a single letter is set italic because it is a variable while a run of letters is a name. Operators and relations get TeX's own spacing, so $\sigma = 1$ reads as an equation rather than as a filename.

A typesetter is installed by wrapping the backend, so it reaches every label the chart has: the title, the axis titles, the ticks, the legend, a facet's strip, a geom's own note. That also means a label is measured as it will be drawn, so the margin left for a fraction is the height of the fraction rather than the width of its markup (ADR 0023). Notation it cannot parse is drawn exactly as written — a chart never fails to render because of a label.

mathtext.Typesetter is the seam if you have a real engine to plug in.

Charts that can be read without being seen

Three channels, because they fail for three different readers (ADR 0024):

p := refract.New(
	refract.Title("Signal against model"),
	refract.Theme(theme.Light.With(theme.Redundant(true))),  // dashes and shapes, not colour alone
)
p.Add(/* ... */)

p.Describe()                      // read the data; write a description
p.Render(refract.SVG("chart.svg"))
p.DataTable(w)                    // the same data as an HTML table

The title alone costs nothing and is always written: an SVG gets <title>, role="img" and aria-labelledby, a PDF gets a document title, a canvas gets role and aria-label. Describe costs a pass over the data — it reports how many rows there are and over what range — so it is a call rather than something every render pays for, and it fills in the <desc> a screen reader announces next:

Signal against model. 3 layers with line marks. Axes: sample horizontally,
σ/√n (mV) vertically. measured, a line of 24 rows, sample from 0 to 23,
measured from 6.16 to 20.9. …

Notation in a title is read aloud rather than spelled out, because "dollar backslash frac" is not a description of anything.

theme.Redundant(true) gives each layer a dash pattern and a marker shape alongside its palette colour — the chart survives a greyscale printout and the readers who cannot separate its first two colours — and it leaves a layer that named its own geom.Dash or geom.Shape alone.

See examples/accessible, which writes the picture, the page and the description as three files.

Charts that follow their surface

p := refract.New(refract.Size(800, 500), refract.Responsive(true))
// ...
live.Resize(400, 250)   // half the size: half the type, half the strokes

A plot is designed at one size and often drawn at another. Responsive scales the theme — type, strokes, spacings, markers, margins — by how much smaller or larger the drawing is, so a chart at a third of its design size is the chart rather than a photograph of it. At the design size the factor is exactly 1, so turning it on cannot change a still you already have (ADR 0025).

Live.Resize is what a window's resize event and a reflowed canvas call. The scales keep whatever they were zoomed to: a reader who dragged a view into place has not asked to leave it.

For a still at another size — a thumbnail of a chart designed larger — name the design explicitly:

refract.New(refract.Size(200, 125), refract.ResponsiveFrom(800, 500))

Nanoseconds at any zoom

A Unix nanosecond count in this century needs 61 bits, and a float64 has 53. Two instants a nanosecond apart are therefore the same number, and an axis zoomed to a microsecond window has nothing left to separate them with.

p.X(scale.Time(scale.Origin(runStart)))   // the domain is nanoseconds since runStart

With an origin near the data, the subtraction happens in int64 and the axis keeps whole nanoseconds for the hundred days either side of it that a float64 counts exactly. A geom reading a time column goes through the axis's own space, so nothing needs converting by hand, and the JSON spec carries the origin so a document reads back as the same axis.

Live data

A data.Stream is appended to from one goroutine and frozen for the renderer on another. It is deliberately not a Source: a table being appended to between two column reads is a table that disagrees with itself.

st := data.NewStream("t", "y").Window(2000)
p.Add(geom.Line(st.Source(), geom.X("t"), geom.Y("y")))

go func() {
	for s := range samples {
		st.Append(scale.Nanos(s.At), s.Value)   // any goroutine, any time
	}
}()

for range ticker.C {
	st.Snapshot()   // freeze what has arrived
	live.Draw()     // draw the frozen view
}

Each Draw compares the frame with the last one and repaints only where they differ; a frame identical to the last is not painted at all. A backend says it can do that by implementing ir.Partial, and one that cannot gets the whole frame as before (ADR 0016). Appending a row and freezing a view both allocate nothing in the steady state, and the benchmark gate keeps it that way.

See examples/stream.

A chart as JSON

doc, err := p.MarshalJSON()      // indented; json.Marshal(p) compacts it
q, err := refract.ParseJSON(doc) // and reads back as the same chart

The document is Vega-Lite-shaped: data.values, mark.type, encoding.x.field, scale.type, facet and resolve mean what they mean in Vega-Lite, so anyone who knows that vocabulary can read one. It is not a Vega-Lite subset and does not claim to be — refract has marks and options Vega-Lite has no name for, and naming them plainly beats smuggling them through a borrowed name. What is guaranteed is the round trip through refract, and there is a test per mark and per scale that renders both and compares the primitives (ADR 0014).

{
  "$schema": "https://github.com/timzifer/refract/spec/v1",
  "width": 640,
  "height": 400,
  "title": "Throughput",
  "data": {
    "values": [{"x": 0, "y": 2}],
    "format": {"parse": {"x": "number", "y": "number"}}
  },
  "encoding": {
    "x": {"type": "quantitative", "scale": {"type": "linear", "nice": true}},
    "y": {"type": "quantitative", "scale": {"type": "linear", "nice": true}}
  },
  "layer": [
    {
      "mark": {"type": "line", "color": "#0072b2"},
      "encoding": {"x": {"field": "x"}, "y": {"field": "y"}}
    }
  ],
  "config": {"theme": "light"}
}

(shown with the objects folded up; the real output puts every field on its own line)

Plotting Arrow data

import "github.com/timzifer/refract/arrow/v18"

src := arrow.Source(rec)      // rec is an arrow.Record
p.Add(geom.Line(src, geom.X("t"), geom.Y("p99")))

A float64 column with no nulls is Arrow's own buffer — no copy, no conversion. Everything else (integers, float32, timestamps, dictionary-encoded strings) converts once on first use and is cached, so a record with forty columns and a chart that plots two pays for two. An Arrow null becomes NaN, which means the missing-data policy you already set covers it (ADR 0013).

How it fits together

   Your spec  ──►  Model  ──►  IR  ──►  Backend  ──►  output
   ─────────      ─────      ────      ───────       ──────
   geoms          scales     ~8        backend/svg     SVG
   scales         coords     drawing   backend/pdf     PDF
   coords         layout     ops       backend/canvas  browser canvas
   theme          ticks                backend/gg      PNG / JPEG / a surface
   facets         panels               backend/window  a native window

The ir.Backend interface is the seam. Geoms never touch a renderer; a renderer never knows what a scale is. That is what lets refract stand on a young, fast-moving graphics stack without being welded to it — the whole gg adapter is about 300 lines (why that matters).

Things ride on that seam without widening it. A render can be watched, so that a pointer can be told which layer drew what it is over (ADR 0015); two frames can be compared, so that a surface repaints only what moved (ADR 0016); and a backend that can carry words, resize itself, or repaint part of a frame says so through an optional interface — ir.Semantics, ir.Resizer, ir.Partial — rather than through a method every backend would have to implement. No identity channel and no damage channel went into the drawing interface.

The native window is the same argument once more: it is a surface that draws with the raster backend and presents the result, so there is one implementation of every mark and a window shows what a file would (ADR 0021).

Documentation

• CONCEPT.md — the design document: motivation, positioning, architecture, roadmap.
• docs/adr — why the open questions were answered the way they were.
• docs/v1-api-audit.md — every exported identifier with a verdict before the API freeze: freeze, change before v1, or defer.
• docs/chart-types.md — every chart form, what draws it today, and what the missing ones would cost.
• docs/benchmarks.md — the benchmark suite: what each benchmark measures, which numbers are gated, and the latest results.
• pkg.go.dev — the API reference, generated from the doc comments.
• CONTRIBUTING.md — building a five-module repository, how to regenerate golden files and figures, and how a release is tagged.
• SECURITY.md — which versions get fixes, and how to report a vulnerability privately.
• CODE_OF_CONDUCT.md — what participating here looks like.

A note on how this was built

Much of refract was written with an AI assistant — Claude, in Claude Code — in the loop, under human direction and review. The design decisions and the arguments in the ADRs are the ones a human signed off on; a good deal of the typing was not. What makes that workable is the same thing the rest of this README describes: every claim here is held up by a golden file, a test or a benchmark that CI runs on every commit, so the code is checked against the behaviour rather than against a plausible-sounding explanation of it.

License

MIT. The core links nothing; backend/gg links only permissively licensed code (gg is MIT, x/image is BSD-3-Clause). That is a requirement rather than a preference — refract must be embeddable by downstream projects under any license.