WIP: Shortcut for count()

[krlmlr]

This is an attempt to improve the run time of TPCH-21 which uses count() over a large table with many groups. Gives a speedup of up to 3. A slippery slope, but might be worthwhile because the vast majority of summaries are simple counts?

This PR

options(conflicts.policy = list(warn = FALSE))
library(tidyverse)

set.seed(20230529)

get_data <- function(N) {
  distr <- (trunc(rexp(N, 3)) + 1)
  id <- rep(seq_along(distr), distr)
  sec <- as.integer(trunc(runif(length(id), 1, 3)))
  tibble(id, sec)
}

data <- get_data(1e4)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>   0.021   0.001   0.022
data <- get_data(1e5)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>   0.107   0.006   0.113
data <- get_data(1e6)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>   1.418   0.050   1.469
data <- get_data(1e7)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>  13.391   0.367  13.763

get_data(40) |>
  filter(id >= 30)
#> # A tibble: 15 × 2
#>       id   sec
#>    <int> <int>
#>  1    30     2
#>  2    31     2
#>  3    32     1
#>  4    32     1
#>  5    33     2
#>  6    34     2
#>  7    35     2
#>  8    35     2
#>  9    35     1
#> 10    36     2
#> 11    37     2
#> 12    38     2
#> 13    38     1
#> 14    39     1
#> 15    40     2

^{Created on 2023-05-29 with reprex v2.0.2}

v1.1.2

options(conflicts.policy = list(warn = FALSE))
library(tidyverse)

set.seed(20230529)

get_data <- function(N) {
  distr <- (trunc(rexp(N, 3)) + 1)
  id <- rep(seq_along(distr), distr)
  sec <- as.integer(trunc(runif(length(id), 1, 3)))
  tibble(id, sec)
}

data <- get_data(1e4)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>   0.034   0.002   0.036
data <- get_data(1e5)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>   0.276   0.007   0.283
data <- get_data(1e6)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>   3.486   0.085   3.576
data <- get_data(1e7)
system.time(count(data, id, sec))
#>    user  system elapsed 
#>  38.666   0.708  39.530

^{Created on 2023-05-29 with reprex v2.0.2}