Using furrr inside a function call makes it slower (ggplot objects)

[dorwon]

This has been a hard one to make a full reprex for, but I'm hoping someone could explain what is going on or a workaround.

Basically I'm working on code that saves a list of ggplot graphs quickly to disk, ones that are often ~2GB using future_iwalk.

I save each page to a temp directory, then compile the completed pages in the final location.

My issue that I've noticed in practice is that when the code is run inside the function it runs signifigantly slower opposed to when it is written out.

This is an example I've been working on but I can't get the issue to show

library(ggplot2)
library(furrr)
library(dplyr)

plan("multisession")

plot_graph <- function(dt) {
  
  dt |> ggplot(aes(x = x, y = y)) +
    geom_point()
}

save_graphs <- function(list_plots) {
  
  dir_out_tmp <- tempdir()
  filename_out <- "temp"
  
  furrr::future_iwalk(
    list_plots,
    ~withr::with_pdf(
      new = fs::path(dir_out_tmp, paste(filename_out, .y, sep = "-"), ext = "pdf"),
      width = 15,
      height = 8,
      code = plot(.x)
    )
  )
  
  files_temp <- fs::path(dir_out_tmp, paste(filename_out, names(list_plots), sep = "-"), ext = "pdf")
  stopifnot(all(fs::file_exists(files_temp)))
  
  qpdf::pdf_combine(
    input = files_temp,
    output = fs::path(dir_out_tmp, file_name_out, ext = "pdf")
  )
}

a <- 1e3
b <- 1e5

dt <- data.frame(
  id = rep(1:a, each = b),
  x = runif(a*b),
  y = runif(a*b)
)

list_plots <- dt |>
  split(f = "id") |>
  purrr::map(plot_graph)


# Run Outside of Function -------------------------------------------------

tictoc::tic()
dir_out_tmp <- tempdir()
filename_out <- "temp"

furrr::future_iwalk(
  list_plots,
  ~withr::with_pdf(
    new = fs::path(dir_out_tmp, paste(filename_out, .y, sep = "-"), ext = "pdf"),
    width = 15,
    height = 8,
    code = plot(.x)
  )
)

files_temp <- fs::path(dir_out_tmp, paste(filename_out, names(list_plots), sep = "-"), ext = "pdf")
stopifnot(all(fs::file_exists(files_temp)))

qpdf::pdf_combine(
  input = files_temp,
  output = fs::path(dir_out_tmp, filename_out, ext = "pdf")
)
tictoc::toc()

# Run in Function ---------------------------------------------------------

tictoc::tic()
save_graphs(list_plots)
tictoc::toc()

The issue does not seem to show up for simple graphs, but in practice for more complicated graphs code this process can take 8 minutes outside of the function but 4 hours when the function is called.

By more complicated graph I mean I notice that my other graph uses scale_fill_manual() and scale_shape_manual() and when I remove them and retest the difference between the in and out of function version decreases