point_interval is inefficient compared to summarise

[lukaseamus]

I routinely run my models with 8 chains and 1e4 iterations per chain. For continuous predictions I then compute predictive distributions for each predictor value in a sequence of 100–200. This is often necessary to correctly represent nonlinear predictions.

I would like to use the point_interval functions to summarise my predictions, but have run into performance issues. Once the number of groups exceed a certain limit, point_interval crashes R. This is not the case for a summarise call that uses the interval functions and yields exactly the same output.

For example,

prediction %>%
  group_by(group1, group2, group3) %>%
  median_qi(var1, var2, var3, .width = c(.5, .8, .9))

fails where

prediction %>%
  group_by(group1, group2, group3) %>%
  summarise(
    across(
      c(var1, var2, var3),
      list(median = median, 
            lower_0.5 = ~ qi(.x, .width = .5)[1],
            upper_0.5 = ~ qi(.x, .width = .5)[2],
            lower_0.8 = ~ qi(.x, .width = .8)[1],
            upper_0.8 = ~ qi(.x, .width = .8)[2],
            lower_0.9 = ~ qi(.x, .width = .9)[1],
            upper_0.9 = ~ qi(.x, .width = .9)[2]),
      .names = "{.col}.{.fn}"
    )
  ) %>%
  ungroup() %>%
  rename(var1 = var1.median, var2 = var2.median, var3 = var3.median) %>%
  pivot_longer(cols = contains("lower") | contains("upper")) %>%
  separate(col = name, into = c("name", ".width"), sep = "_(?=[^_]*$)") %>%
  pivot_wider(names_from = name, values_from = value)

doesn't.

I really appreciate the succinctness of the point_interval functions, but find myself having to resort to my own summarise functions. Perhaps you could rethink this family of functions to make them as efficient as summarise?