Benchmark

The output of versus::compare() is designed to contain only the minimal amount of information needed to represent the difference between two data frames: rows with value differences, unmatched columns, and unmatched rows. While other packages offer much more functionality, the minimal approach makes versus much faster.

Below is an example benchmark using data from the nycflights13 package over-sampled to various sizes. In the example data, 5% of rows from each data frame are missing in the other, and in 4 out of the 15 columns, 5% of values are different.

At 1 million rows, versus takes under half a second and half a gigabyte of memory, while others take over 10 seconds and multiple gigabytes of memory. This is made possible by efficient functions from vctrs and collapse. The compare() operation is roughly described by the steps below.

• Locate matching rows between table_a and table_b with vctrs::vec_locate_matches
• Subset tables according to match indices using collapse::ss
• Use collapse::%!=% to identify differing values

Benchmark code:

suppressPackageStartupMessages(library(dplyr))

generate_example_data <- function(n_rows) {
  tbl <- nycflights13::weather |>
    slice_sample(n = n_rows, replace = TRUE) |>
    # simulate unique key
    mutate(time_hour = Sys.time() + row_number())

  tbl_a <- tbl |>
    slice_sample(n = floor(n_rows * 0.95))
  tbl_b <- tbl |>
    slice_sample(n = floor(n_rows * 0.95)) |>
    # simulate value differences
    mutate(across(temp:wind_dir, \(x) ifelse(runif(n()) < .05, x + 1, x)))
  list(a = tbl_a, b = tbl_b)
}

bench_out <- bench::press(
  n_rows = c(1e5, 1e6, 2e6),
  {
    tbl <- generate_example_data(n_rows)
    bench::mark(
      versus =
        versus::compare(tbl$a, tbl$b, by = c(origin, time_hour)),
      arsenal =
        arsenal::comparedf(tbl$a, tbl$b, by = c("origin", "time_hour")),
      dataCompareR =
        dataCompareR::rCompare(tbl$a, tbl$b, keys = c("origin", "time_hour")),
      diffdf =
        diffdf::diffdf(tbl$a, tbl$b, keys = c("origin", "time_hour"), suppress_warnings = TRUE),
      min_iterations = 2,
      max_iterations = 2,
      check = FALSE
    )
  }
)

Benchmark results:

# benchmark run on 2020 i7 MBP with 32GB of memory
bench_out |>
  summary() |>
  mutate(across(n_rows, scales::comma)) |>
  select(1:6) |>
  print(n = Inf)
#> # A tibble: 12 × 6
#>    expression   n_rows         min   median `itr/sec` mem_alloc
#>    <bch:expr>   <chr>     <bch:tm> <bch:tm>     <dbl> <bch:byt>
#>  1 versus       100,000    41.42ms  61.31ms  16.3       40.08MB
#>  2 arsenal      100,000   641.63ms 668.55ms   1.50        389MB
#>  3 dataCompareR 100,000      1.09s    1.18s   0.848    265.39MB
#>  4 diffdf       100,000      3.45s    3.54s   0.283    554.95MB
#>  5 versus       1,000,000 383.67ms 386.76ms   2.59     403.69MB
#>  6 arsenal      1,000,000   11.63s   11.67s   0.0857     3.77GB
#>  7 dataCompareR 1,000,000   15.99s   16.44s   0.0608     2.63GB
#>  8 diffdf       1,000,000   54.96s   55.04s   0.0182     5.35GB
#>  9 versus       2,000,000    1.16s     1.2s   0.833    806.13MB
#> 10 arsenal      2,000,000   24.05s   26.24s   0.0381     7.54GB
#> 11 dataCompareR 2,000,000   38.77s   39.85s   0.0251     5.27GB
#> 12 diffdf       2,000,000    1.89m    1.97m   0.00847    10.7GB