pipes vs data.table

pipes_brackets.R

library(data.table)
library(magrittr)
library(ggplot2)
set.seed(42)
do_bench <- function(N) {
  dt <- suppressWarnings(data.table(x = rnorm(N), 
                   group = letters[1:3]))
  
  pipe <- function() {
    dt %>% 
      .[, mean(x), by = group] %>% 
      .[group == "a"]
  }
  
  
  bracket <- function() {
    dt[, mean(x), by = group][group == "a"]
  }
  
  
  
  bm <- bench::mark(pipe, bracket, check = FALSE)
  
  as.numeric(bm$time[[2]]/bm$time[[1]])
}

benches <- data.table(N = 10^seq(4, 8)) %>% 
  .[, do_bench(N), by = N] 
  

ggplot(benches, aes(factor(N), V1)) +
  geom_boxplot() +
  ggforce::geom_sina(size = 0.1, alpha = 0.1) +
  geom_hline(yintercept = 1) +
  scale_y_log10("Timing [][] / %>%", breaks = c(1/3, 1/2, seq(1, 3)), limits = c(1/4, 4),
                labels = c("1/3", "1/2", seq(1, 3))) + 
  scale_x_discrete("Number of rows") +
  labs(title = "Using pipes with data.table operations has negligible impact on timing",
       subtitle = "doing \ndt %>% .[, mean(x), by = group] %>% .[group == 1] vs. \ndt[, mean(x), by = group][group == 1]") +
  hrbrthemes::theme_ipsum_rc()