NEWS.md
In glyph: A Grammar of Interactive Graphics

glyph 0.2.0

scale() is now set_scale() (#36). glyph's scale() had the same name as base::scale(), so library(glyph) masked the base function. It printed "The following object is masked from 'package:base': scale", and base uses such as scale(df$x) failed afterwards with an error about glyph specs. set_scale() takes the same arguments. To update your code, change each scale() step in a glyph pipeline to set_scale():

r # Before glyph(mtcars, x = wt, y = mpg) |> mark_point() |> scale("y", zero = TRUE) # After glyph(mtcars, x = wt, y = mpg) |> mark_point() |> set_scale("y", zero = TRUE)

Base uses of scale() work again and need no change. scale_color(), scale_log() and scale_time() keep their names. library(glyph) no longer masks any function. The compiled JSON (compile()$json and export(..., "plot.json")) has a new data layout, marked by "$schema": "glyph/v0.2" (#8). Data is {"columns": {...}, "n": rows}, and a mark that draws the plot's data has {"from": "plot", "fields": [...], "rows": [...]}, where rows is null or lists the 0-based rows it keeps after dropping missing values. Boxplot, violin, histogram and polygon marks send {"values": [...]}, one record per box, violin, bin or shape. Code that reads JSON written by glyph 0.1 needs updating. compile(engine = ...) accepts only "auto" and "html", and gives an error for anything else. It used to accept "canvas" and "webgl" and pick one by data size, but no such renderer has ever existed, so the choice did nothing. The D3 and SVG renderer is the only one. * Printing a compose() layout at the console now draws it, as printing a single plot does (#13). The text description it used to print comes from the new summary.glyph_layout().

glyph works in Shiny (#34). glyphOutput() and renderGlyph() place and draw a plot, and renderGlyph() takes a plot, a compose() layout or a widget from render(). What the user selects reaches the server as input$<outputId>_selected: the row numbers of the data frame given to glyph(), so data[input$plot_selected, ] gives the selected rows. The numbers stay right when rows with missing values are dropped and when a mapping is computed, and selecting a box, bin or polygon selects the rows it summarises. The value is sent when a brush ends, on a click select and on a legend click, and clearing a selection sends NULL. A facet() grid gives one vector, and a compose() layout gives a list with one entry per panel, named after the panels when compose() is given names. shiny is suggested, not required. A new article, "Using glyph in Shiny" (vignette("shiny", package = "glyph")), is built around an example app in inst/examples/shiny-brush.
New statistical marks (#6): mark_boxplot(), mark_violin(), mark_sina() and mark_histogram(). Their statistics are computed in R when the plot is compiled, with stats::quantile(), stats::density() and hist()-style bins, after computed mappings are evaluated and rows with missing values are dropped. They work with facet() (summaries are computed per panel), compose(), theme tokens, tooltips (which show the statistics), hover, click and brushing. Each box, violin or bin knows which rows it summarises, so brushing points or linked panels highlights the boxes and bins that contain the selected rows, and brushing a box selects its rows.
New mark_polygon() draws closed shapes from x/y vertices, one per value of group (#7). Vertices are joined in row order, color fills each shape, style sets the fill, outline colour, outline width and opacity, and radius rounds the corners. Each polygon is one element: it gets one tooltip, and brushing or clicking it selects all of its rows, including in linked panels. It works with facet(), compose(), zoom, theme tokens, computed mappings and NA dropping.
Plots and layouts now draw automatically in R Markdown, Quarto and vignettes, without ending the pipeline with render() (#13). glyph registers knit_print() methods for glyph_spec and glyph_layout; knitr stays a suggested package. render() is still there for setting the width and height.
Plots with many rows are much quicker to prepare (#8). Plot data now goes to the browser as one array per column instead of one object per row, and is turned back into rows in the browser before anything is drawn. Each row is sent once: a mark that draws the plot's own data refers to it instead of carrying a second copy. Boxplot, violin and histogram marks send their summaries and the row indices that brushing and linked selections use, not the raw rows. At 100,000 points compile() takes 0.1s instead of 65s, the widget is ready for the Viewer in 0.4s, and the file is 5 MB instead of 14 MB. Plots look and behave as before.
New "Clear selection" button in a plot's controls bar, shown while a brush selection exists (#26). Clicking or tapping it clears the selection and restores every mark, including marks in panels linked with compose(linked_selections = TRUE). Pressing Escape clears brush selections too, unless you are typing in a text field. Plots with a brush but no zoom now get a controls bar for the button. When the controls don't fit the plot's width, as in small facet panels, they show just their symbols, with the full label in their tooltip.
New reset-zoom control. A double tap now resets the zoom instead of zooming in further, so a zoomed plot can be zoomed back out on a touch device.
Tooltips now appear on tap, and a tap elsewhere dismisses them, for touch devices that have no hover.
New article, "Adding a custom mark" (vignette("custom-marks")), walks through adding a mark type end to end with a worked example (#9). It covers the R function and add_mark(), how mappings are resolved when the plot is compiled, how data reaches the browser, the drawing code in glyph.js, axes, tooltips, hover, brushing, fitting the container, tests, and marks that summarise data. It also says plainly that adding a mark means editing glyph's source, and what would need to change for other packages to add marks.
Rewrote the README, the pkgdown home page, ARCHITECTURE.md and the Getting Started vignette in plainer language (#4). Removed the comparisons with ggplot2 from the documentation, including the comparison tables in the README and the "glyph vs ggplot2" vignette. The package Title and Description now describe what glyph does (#5). The README and the glyph-package help page have an "Acknowledgements" section crediting the prior art glyph's design draws on.
For contributors: browser tests (#28) draw plots in headless Chrome through chromote and check what was drawn. They check several of the bugs fixed in this release, and the selections sent to Shiny. They are skipped on CRAN and wherever chromote or Chrome isn't available. GitHub Actions runs them, the pkgdown site build, and R CMD check on Windows, macOS and Linux for every pull request (#31).

Bar and area charts now start their value axis at zero (#29). The y axis used to start at the smallest value, so the smallest bar had no height and the others weren't in proportion to their values. Bars and areas are now drawn from zero, so a negative bar hangs down from zero instead of rising from the bottom of the plot. Histograms already started at zero. A domain set with set_scale() is used as given, including on the shared axes of facet() and compose(). Point and line marks are unchanged.
Plots now fit inside their container in both directions (#17). A plot was drawn at the spec's width and height (600 x 400 unless set) and scaled to the container's width, so in a wide, short container such as RStudio's Viewer pane it came out taller than the space available and its x axis was cut off. The same happened to facet() grids, which also overflowed narrow containers, and plots in R Markdown documents ran past their 400px box into the text below. Plots and facet grids are now drawn at the size of the container they are in, and are redrawn when it is resized.
Tooltips no longer fall behind the pointer over dense areas of a large point plot (#23). With brushing on, each pointer move used to search every point for the nearest one and move the tooltip, and in WebKit moving the tooltip repainted the plot underneath it. At 15,000 points one move took about half a second, so moves queued up and the tooltip kept changing for over a minute after the pointer stopped. Points are now found through a quadtree, pointer moves are handled at most once per frame for the latest position, and the tooltip is drawn on its own layer so moving it doesn't repaint the plot. The same points are found as before.
Redrawing a plot no longer leaves old tooltips in the page (#23). Every resize redraws the plot, and fitting it to its container can take up to three draws. Each draw added another tooltip for every mark with one, plus click listeners that kept the replaced plot in memory. A faceted plot resized ten times left 144 tooltips behind, and a tooltip showing when the plot was redrawn could stay on screen. Each widget now has one tooltip, which a redraw hides and reuses.
"Reset zoom" and double-click zooming no longer lag on large plots (#25). They jump straight to the result instead of animating when the plot has 3,000 or more elements, since each frame of the animation redrew every point. At 15,000 points a reset now takes about 50ms instead of 470ms. Each zoom step still redraws every point, which canvas drawing (#20) is meant to fix.
compose() with named arguments, such as compose(a = p1, b = p2), now draws (#34). The panels were sent as an object keyed by name instead of a list, and the widget showed a render error.
export() to .html no longer fails on machines without pandoc (#3). Instead of a low-level pandoc error, it now writes the HTML with its JavaScript in a <name>_files folder alongside, and says so, including how to get a single self-contained file (install pandoc, or run inside RStudio). The pandoc requirement is now documented in ?export.
The entrance (slide) animation for bar charts no longer shows no bars on Safari and Chrome mobile (#2). The animation read a value that was never set, so bars always animated to zero height. It now stores each bar's real height and position before animating.
Brushing and zooming now apply to the whole plot rather than to each point mark separately (#6). A plot with several brushable marks gets one brush that covers all of them, and interact(zoom = TRUE) now zooms plots that have no point mark.
Points on a categorical x axis are now drawn in the middle of their category instead of at its left edge (#6).
A quantitative axis now spans the values of every mark in the plot, not just the first, so a second mark with a wider range is no longer cut off (#6).
Automatic tooltips show a computed mapping by its expression (such as factor(cyl)) instead of its internal column name, and tooltip templates can refer to fields whose names contain dots, such as {.median} (#6).
Boxplot, violin, sina and histogram marks no longer leave out rows whose colour is missing. Those rows now form a group of their own (#7).
Computed aesthetic mappings, such as x = log(wt) or color = factor(cyl), are now evaluated when the plot is compiled. They used to be dropped without a message. The expression is used as the default axis or legend label.
Rows with a missing value in a mapped position aesthetic (x, y, x2, y2, y_min, y_max) are now dropped from that mark when the plot is compiled, with a warning that says how many rows were removed and for which aesthetics.
facet(rows =, cols =) now accepts a computed expression, such as facet(cols = factor(cyl)), evaluated the same way as an aesthetic mapping. The expression text used to be taken as a column name that didn't exist, so faceting on anything but a plain column gave a single, unlabelled panel.
Faceted panels now use each mark's own resolved data, with its computed columns and with rows dropped for missing values, instead of an unfiltered slice of the shared data. So facet() with a computed mapping or with missing values behaves the same as an unfaceted plot.
Plots scale to fill their container instead of being drawn 600px wide, and the theme background always covers the whole plot area. A dark theme's background used to stop short of the edges in narrower or wider containers, including on phones.
With brushing on, tapping or clicking a point now reaches it, so its tooltip and click select work. The brush's overlay covers the plot so a drag can start anywhere, and it used to swallow the tap. Hovering over a point with brushing on now shows its tooltip and hover effect too, for the same reason. Brushing still works by dragging from anywhere in the plot.
With both zoom and brushing on, one drag no longer both pans and brushes. A drag now does one thing at a time. It brushes by default, and a small toggle next to the reset-zoom control switches dragging between selecting and panning. The scroll wheel, a trackpad pinch, a two-finger pinch and a double-click or double tap always zoom, whichever mode the toggle is in.
A two-finger pinch now zooms plots that also have brushing on. d3-brush stopped every touch move from reaching the zoom once its one-finger gesture started. The pinch is now handled first.
The plot no longer jumps by its axis margin the first time it is zoomed or panned. The zoom transform replaced the margin offset instead of adding to it.
Clicking the reset-zoom control no longer clears the brush selection.
A brush selection now also clears on a tap outside the plot, or on the plot's margin, not only on a tap inside the plot area.
A double tap zooms or resets reliably, including repeated double taps and while dragging is set to brushing. It no longer depends on d3-zoom's own gesture detection.
The help pages no longer compare glyph with other packages, and several describe the code more accurately (#5). ?glyph no longer claims that lists, URLs or Arrow tables are accepted as data, ?facet no longer says panels get independent scales by default, and ?marks and ?interactivity drop features that don't exist. The description, documentation, README and ARCHITECTURE.md no longer claim WebGL or canvas rendering or standalone SVG export. A canvas renderer is planned (#20).
?mark_text now describes what smart_repel = TRUE does: it nudges overlapping label boxes apart, pair by pair, over several passes. It credits ggrepel as the inspiration.
The summary.glyph_spec and print.glyph_layout help topics were left out because they had no title. They are back.
Mike Bostock is now credited in Authors@R and inst/COPYRIGHTS as the author of the bundled D3.js library (ISC license), as CRAN policy asks for bundled JavaScript.

glyph 0.1.1

Fixed critical rendering bug where interactive plots displayed as blank in RStudio viewer, R Markdown documents, and Shiny apps. Data was being serialized in column-oriented format instead of the row-records format expected by the 'D3.js' frontend.
Added error surfacing in the htmlwidget so JavaScript failures display a visible error message instead of a silent blank div.
Added sizing policy for better default widget dimensions in R Markdown and pkgdown contexts.

glyph 0.1.0

Initial CRAN release.
Core spec builder with quosure-based aesthetic mappings (no aes() required).
Eight mark types: mark_point(), mark_line(), mark_bar(), mark_area(), mark_text(), mark_rule(), mark_ribbon(), mark_link().
First-class interactivity via interact(): tooltips, zoom, brush selection, hover effects, cross-filtering, and named selections.
Declarative animation via animate(): entrance transitions and keyframe animation through data states.
Unified scale system via scale() with convenience wrappers scale_color(), scale_log(), scale_time().
Token-based theming via theme_tokens() with five built-in presets: light, dark, minimal, publication, presentation.
Built-in layout composition via compose(), marginals(), inset(), and facet().
Compilation pipeline via compile(), rendering to interactive HTML widgets via D3.js.
JSON spec export via export() and Vega-Lite interoperability via to_vegalite().
D3.js v7 rendering engine via 'htmlwidgets'.