In daattali/rsalad: A mix of useful R functions that are good for you

knitr::opts_chunk$set(tidy = FALSE, comment = "#>")

Performance of `dfCount()`

After some basic testing (not extremely thorough), I believe that dfCount() performs much faster than its equivalent table() on large datasets, especially when the data is numeric. The analysis was done with the microbenchmark package to compare the two functions on a few different datasets.

library(rsalad)
library(dplyr)
library(microbenchmark)

# Prepare all the datasets to test on
fDat <- nycflights13::flights
largeIntDat <- data.frame(col = rep(1:25, 100000))
largeCharDat <- data.frame(col = rep(letters[1:25], 100000))
smallDat <- data.frame(col = rep(1:25, 100))

# Run the benchmarking
m <-
  microbenchmark(
    dfCount(fDat, "day"), table(fDat$day),
    dfCount(fDat, "dest"), table(fDat$dest),
    dfCount(largeIntDat, "col"), table(largeIntDat$col),
    dfCount(largeCharDat, "col"), table(largeCharDat$col),
    dfCount(smallDat, "col"), table(smallDat$col),
    times = 10
  )

knitr::kable(summary(m) %>% select(expr, min, mean, median, max, neval))

Every pair of rows corresponds to counting the same data using dfCount() vs table(). The results show that:

dfCount() was faster in all 4 large datasets
dfCount() was an order of magnitude faster in both cases when the data was numeric
dfCount() was slower on very the small dataset

After performing this analysis, I've realized that the likely cause of the speed boost is due to dfCount() relying on dplyr. After making that realization, I found that dplyr also has a count() function, which performs equally fast as dfCount(), which further supports the hypothesis that the speed boost was thanks to dplyr. However, I still want to include this function in the package because it took a lot of hard work (and documentation!), and it also has a very differences from dplyr::count(). For example, dplyr::count() does not sort by default, which I find to be the less desired behaviour, and dplyr::count() does not have a standard-evaluation version.