A big scatter that stays interactive

120,000 points you can still hover, brush, and zoom. Past 20,000 keyed marks, as_widget() draws the scene once as an image and runs every interaction off a compact index.

mark_datashade() handles a million points by binning them into a grid and shading each cell by how many points land in it. You get the shape of the cloud, but the individual point is gone, so there is nothing to hover and no way to brush a few of them out.

as_widget() takes a different route that keeps every point addressable. Once a plot passes 20,000 keyed elements it stops writing one SVG node per point. A 150,000-point scatter used to ship a 75 MB file with 150,000 DOM nodes and stutter on every hover. Now the scene is drawn once as a single embedded image, and hover, brush, and pan/zoom run off a compact index of bounding boxes and keys. The page holds a picture and a lookup table in place of the nodes.

Here are 120,000 points in six colour groups, each spread across a few tight sub-blobs. Hover a point for its label, drag to brush a region, scroll to zoom, and click a legend swatch to isolate one group. It is a static HTML file with no server behind it.

set.seed(42)
groups <- c("Alpha", "Beta", "Gamma", "Delta", "Epsilon", "Zeta")

per_group <- 20000L

make_group <- function(gi) {
  nb <- 7L
  bx <- runif(nb, -9, 9)
  by <- runif(nb, -9, 9)
  b <- sample(nb, per_group, replace = TRUE)
  data.frame(
    x = bx[b] + rnorm(per_group, 0, 0.55),
    y = by[b] + rnorm(per_group, 0, 0.55),
    group = groups[gi]
  )
}

df <- do.call(rbind, lapply(seq_along(groups), make_group))
df$group <- factor(df$group, levels = groups)
df$id <- seq_len(nrow(df))

p <- vplot(df) |>
  mark_point(
    x = x, y = y, color = group, data_id = id,
    tooltip = paste0(group, " · point #", id), size = 0.6
  ) |>
  labs(title = "120,000 points, still interactive")

as_widget(p, height = 820)

When the image path kicks in

The switch is automatic. The default mode = "auto" counts the keyed elements and moves to the image path above raster_threshold, which is 20,000; anything smaller keeps the per-element SVG it always used. Pass mode = "svg" or mode = "raster" to force one path.

Zooming stays sharp. The base image is anti-aliased for the full view, and when you zoom in the widget redraws just the visible points on a canvas overlay, reading each point’s colour from the rendered image and its position from the index. A zoomed region resolves into crisp points rather than a blurry upscale.

What you give up: with no node per point, per-element grammar colours and per-mark screen-reader focus no longer apply, and crosstalk’s cross-filter is off. The trade is a fast, navigable cloud in place of 120,000 live DOM elements.