Run plot comms on the R thread

[lionel-]

Progress towards #689
Branched from #1099

I did not foresee all the simplifications falling out of this one!

• Plots no longer live in their own threads, they live on the R thread. That frees up memory since each background thread takes about 2mb of stack space (depending on platform), so 20 open plots would take 40mb of memory.

• DeviceContext moves from a standalone thread_local! into a regular field of Console. All access goes through Console::get().device_context(). The CommHandlerContext provides access to Console via ctx.console(), so plot handlers reach the device context through a controlled context chain.

• The old idle-time polling loop (process_rpc_requests with crossbeam::Select) is removed. Plot RPCs are now dispatched synchronously through the shell handler on the R thread. This improves plot latency when pre-renderings require adjustments because we no longer timeout-poll for plot updates (related to Performance for plotting positron#5184).

• The GraphicsDeviceNotification async channel is removed. Since the UI comm now runs on the R thread (from Run UI comm on the R thread #1099), prerender settings are updated synchronously via Console::get().device_context().set_prerender_settings(). This removes quite a bit of plumbing.

• Fixes a dev.hold() regression I introduced with pre-renderings (Send pre-renderings of new plots to the frontend #775). We were previously sending pre-renderings unconditionally, now they are held until dev.flush(). This is tested by new integration tests.

[DavisVaughan]

I have paused on review until we can get these regressions resolved

[DavisVaughan]

This smells a bit funny. I don't think you should need this if things are working correctly

[DavisVaughan]

It makes sense to me to remove iopub_tx from DeviceContext now and only pull iopub_tx directly from Console

Over in graphics_device.rs, you can just Console::get().iopub_tx() anytime you need it now!

It feels like this better scopes responsibilities to me

[lionel-]

I feel the opposite. It hides that the device context directly communicates with the frontend via IOPub.

Also ideally we should avoid Console::get() calls as these bypass Rust's ownership model. But then the issue is that the device belongs to console and not the other way around. Cloning the IOPub channel is the right ownership structure from that point of view.

[DavisVaughan]

Is it possible that we could now make DeviceContext have non-Cell internals?

I am disturbed by being able to call get() to then perform a mutable update, rather than being forced to go through get_mut()

[DavisVaughan]

I am pretty sure these are important

[DavisVaughan]

Broken regression from this PR

plt2 = function() {
  layout(matrix(1:2, 2))
  plot(1, 1)
  plot(1, 1)
}
plt2()

[DavisVaughan]

In general, just send this whole chunk "all at once" rather than in separate requests

  layout(matrix(1:2, 2))
  plot(1, 1)
  plot(1, 1)

[DavisVaughan]

Broken regression

# Stars raster
library(stars)
library(terra)
r <- matrix(runif(9), nrow = 3) |>
  st_as_stars()
plot(r)
rast(r) |> plot()

[DavisVaughan]

Broken regression

# rpart
library(rpart)
library(rpart.plot)
model <- rpart(Species ~ ., data = iris, method = "class")
rpart.plot(model)

[DavisVaughan]

Actually something somehow seems deeply broken by this? I am getting blank new-pages fairly often and reproducibly. Run these 3 plots one at a time.

# rpart
library(rpart)
library(rpart.plot)
model <- rpart(Species ~ ., data = iris, method = "class")
rpart.plot(model)

# sf
geo <- sf::st_make_grid(
  cellsize = c(1, 1),
  offset = c(0, 0),
  n = 10
)
x <- sample(c(0, 1), 100, prob = c(0.9, 0.1), replace = TRUE)
plot(sf::st_sf(x = x, geo))

# rpart again
library(rpart)
library(rpart.plot)
model <- rpart(Species ~ ., data = iris, method = "class")
rpart.plot(model)

Screen.Recording.2026-03-16.at.4.34.47.PM.mov

[DavisVaughan]

In general here is my "regression script" I run through one at a time (following the per example instructions of whether to run it one line at a time or as a chunk of code)

All of these should work

# -------------------------
# Multiple plots at once

# Run the chunk all at once
plot(1:10)
plot(2:20)

# Run the chunk all at once
library(ggplot2)
g <- ggplot(mpg, aes(class))
# Number of cars in each class:
g + geom_bar()
# Total engine displacement of each class
g + geom_bar(aes(weight = displ))
# Map class to y instead to flip the orientation
ggplot(mpg) + geom_bar(aes(y = class))

# Inside a function
plt2 = function() {
  layout(matrix(1:2, 2))
  plot(1, 1)
  plot(1, 1)
}
plt2()

# Multiple ggplots
library(ggplot2)
library(purrr)
economics_list <- split(economics_long, economics_long$variable)
economics_list |>
  imap(\(df, nm) {
    df |>
      ggplot(aes(date, value)) +
      geom_line() +
      labs(title = nm)
  })

library(ggExtra)
library(ggplot2)
set.seed(30)
df <- data.frame(x = rnorm(500, 50, 10), y = runif(500, 0, 50))
p2 <- ggplot(df, aes(x, y)) + geom_point()
ggMarginal(p2, type = "boxplot")
ggMarginal(p2, type = "histogram")
par(mfrow = c(1, 1))

# -------------------------
# Plotting in a loop

# (Note it would be nice if we could process renders at interrupt time)

for (i in 1:5) {
  plot(i)
  Sys.sleep(1)
}

# -------------------------
# Par related examples

par(mfrow = c(2, 2))
hist(
  iris$Petal.Width[iris$Species == "virginica"],
  xlim = c(0, 3),
  breaks = 9,
  main = "Petal Width for Virginica",
  xlab = "",
  col = "green"
)
hist(
  iris$Petal.Width[iris$Species == "versicolor"],
  xlim = c(0, 3),
  breaks = 9,
  main = "Petal Width for Versicolor",
  xlab = "",
  col = "blue"
)
hist(
  iris$Petal.Width[iris$Species == "versicolor"],
  xlim = c(0, 3),
  breaks = 9,
  main = "Petal Width for Versicolor",
  xlab = "",
  col = "blue"
)
plot(1:10)
par(mfrow = c(1, 1))


par(mfrow = c(2, 1))
plot(rnorm(100))
plot(rnorm(1000))
par(mfrow = c(1, 1))

# Stars raster
library(stars)
library(terra)
r <- matrix(runif(9), nrow = 3) |>
  st_as_stars()
plot(r)
rast(r) |> plot()

# Performance models
library(performance)
data(mtcars)
m <- lm(mpg ~ gear + cyl + hp, data = mtcars)
check_model(m)

m <- lm(mpg ~ gear + cyl + hp, data = mtcars)
p <- plot(check_model(m), panel = FALSE)
p[[1]] # works
p[[2]] # works
# blank plot
patchwork::wrap_plots(p)

# rpart
library(rpart)
library(rpart.plot)
model <- rpart(Species ~ ., data = iris, method = "class")
rpart.plot(model)

# layout() related
# layout(matrix(
#   c(1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2),
#   nrow = 4,
#   byrow = TRUE
# ))
# layout.show(n = 2)

# sf
geo <- sf::st_make_grid(
  cellsize = c(1, 1),
  offset = c(0, 0),
  n = 10
)
x <- sample(c(0, 1), 100, prob = c(0.9, 0.1), replace = TRUE)
plot(sf::st_sf(x = x, geo))

# biplot
dat <- mtcars
pr <- princomp(dat)
biplot(pr)

# -------------------------
# dev.hold()-ing

# Run these one at a time!
# (RStudio can't actually do this)
dev.hold()
plot(1:10)
abline(1, 2)
dev.flush()

# -------------------------
# tinyplots

library(tinyplot)

plt(Sepal.Length ~ Petal.Length, data = iris)
plt(Sepal.Length ~ Petal.Length | Species, data = iris)
plt(Sepal.Length ~ Petal.Length | Species, data = iris, legend = "bottomright")

# generate data
lambda1 <- 1.8
lambda2 <- 0.7
x <- 0:10
p1 <- dpois(x, lambda1)
p2 <- dpois(x, lambda2)
dp <- data.frame(
  x = c(x, x),
  p = c(p1, p2),
  lambda = rep(c(lambda1, lambda2), each = length(x))
)
dp$lambda <- as.factor(dp$lambda)
# plot
plt(
  p ~ x,
  facet = ~lambda,
  data = dp,
  type = "h",
  lwd = 2,
  col = "skyblue",
  draw = c(
    axis(1, at = x, labels = x),
    points(p ~ x, data = dp, col = "black")
  )
)

# -------------------------
# The dreaded graphics device swap

# Run this all at once
library(ggplot2)
p <- ggplot(mtcars, aes(x = wt, y = mpg)) + geom_point()
p
ggsave("temp.png", p)

# Run this all at once
plot(1:10)
grDevices::png()
plot(1:20)
dev.off()

# -------------------------
# Something related to big plots was not working when you resize
# (I did some perf improvements)

library(qgraph)
data(big5)
data(big5groups)
# First plot
qgraph(
  cor(big5),
  minimum = 0.25,
  groups = big5groups,
  legend = TRUE,
  borders = FALSE,
  title = "Big 5 correlations"
)
# Second identical plot
qgraph(
  cor(big5),
  minimum = 0.25,
  groups = big5groups,
  legend = TRUE,
  borders = FALSE,
  title = "Big 5 correlations"
)

# -------------------------
# Doesn't work - stepping through plots one at a time
# Not precisely sure why, because in theory we are at idle time

fit_lm <- lm(
  mpg ~ cyl + wt,
  data = mtcars
)

# Renders only plot(fit_lm, which = 5)
# Not able to scroll through the which = 1, 2, 3 plots
# that are automatically called by plot.lm().
plot(fit_lm)

# -------------------------
# Doesn't work - we don't respond to render requests in the middle of R execution
# We respond at idle time right now in `process_events()`.
# So all the plots appear at once at the end

tourr::animate_xy(tourr::flea[, 1:6])

# Simpler example of the idea
{
  plot(1:10)
  plot(2:11)
  Sys.sleep(5)
}

[lionel-]

Thanks! I'll add these as test cases