Using `profmem` package to understand data summaries by group for both `data.table` and `dplyr`

dplyr_datatable_memory.R

library(bench)      # assess speed and memory
library(data.table) # data.table for all of its stuff
library(dplyr)      # compare it to data.table
library(profmem)    # assess the process of R functions

set.seed(84322)

# Example Data
d <- data.table(
  grp = sample(c(1,2), size = 1e6, replace = TRUE) %>% factor,
  x = rnorm(1e6),
  y = runif(1e6)
)
d

#>          grp          x         y
#>       1:   1  0.2558379 0.2034364
#>       2:   2 -0.8886153 0.4684875
#>       3:   2  0.4724519 0.6850357
#>       4:   1  0.7360537 0.4890217
#>       5:   1  0.6855063 0.6964860
#>      ---                         
#>  999996:   1  2.1008965 0.3624327
#>  999997:   2  2.2423628 0.2595716
#>  999998:   2  1.5314115 0.1102460
#>  999999:   2 -1.6086973 0.2679477
#> 1000000:   2  1.2518419 0.6566943

# data size and group vector size
lobstr::obj_size(d)
lobstr::obj_size(d$grp)

#> 20,001,752 B
#> 4,000,560 B

# Copy all data for summaries
df <- copy(d) %>% as.data.frame()
tbl <- copy(d) %>% as_tibble()
dt <- copy(d)

# profile dplyr
profmem::profmem(summarize(group_by(tbl, grp), mean(x))) %>% 
  data.frame %>% 
  select(bytes, calls)

#>     bytes                                                                                   calls
#> 1 1997104 summarize() -> group_by() -> group_by.data.frame() -> grouped_df() -> grouped_df_impl()
#> 2 2003000 summarize() -> group_by() -> group_by.data.frame() -> grouped_df() -> grouped_df_impl()
#>

# benchmark dplyr
bench::mark(summarize(group_by(tbl, grp), mean(x)), 
            iterations = 25) %>% 
  select(median, mem_alloc)
  
#> # A tibble: 1 x 2
#>     median mem_alloc
#>   <bch:tm> <bch:byt>
#> 1   26.7ms    3.81MB

# profile data.table
profmem::profmem(dt[, mean(x), by = grp]) %>% 
  data.frame %>% 
  select(bytes, calls)
  
#>       bytes                                                       calls
#> 1       280                          [() -> [.data.table() -> new.env()
#> 2      8240 [() -> [.data.table() -> null.data.table() -> setalloccol()
#> 3      8240 [() -> [.data.table() -> null.data.table() -> setalloccol()
#> 4   4000048                          [() -> [.data.table() -> forderv()
#> 5   2003000                                       [() -> [.data.table()
#> 6      8248                                       [() -> [.data.table()
#> 7   4005944                                       [() -> [.data.table()
#> 8      8248                                       [() -> [.data.table()
#> 9       280                          [() -> [.data.table() -> new.env()
#> 10      248                                       [() -> [.data.table()
#> 11      248                                       [() -> [.data.table()
#> 12      248                                       [() -> [.data.table()
#> 13      248                                       [() -> [.data.table()
#> 14      248                                       [() -> [.data.table()
#> 15      248                                       [() -> [.data.table()
#> 16      248                                       [() -> [.data.table()
#> 17      248                                       [() -> [.data.table()
#> 18      248                                       [() -> [.data.table()
#> 19      248                                       [() -> [.data.table()
#> 20      248                                       [() -> [.data.table()
#> 21      248                                       [() -> [.data.table()
#> 22      248                                       [() -> [.data.table()
#> 23      248                                       [() -> [.data.table()
#> 24      248                                       [() -> [.data.table()
#> 25  4000056                           [() -> [.data.table() -> gforce()
#> 26  2000056                           [() -> [.data.table() -> gforce()
#> 27  2000056                           [() -> [.data.table() -> gforce()
#> 28 16000056                           [() -> [.data.table() -> gforce()
#> 29     8256                      [() -> [.data.table() -> setalloccol()
#> 30     8256                      [() -> [.data.table() -> setalloccol()

# benchmark with gforce
bench::mark(dt[, mean(x), by = grp], 
            iterations = 25) %>% 
  select(median, mem_alloc)
  
#> # A tibble: 1 x 2
#>     median mem_alloc
#>   <bch:tm> <bch:byt>
#> 1   16.8ms    32.5MB

# profile data.table without gforce
profmem::profmem(dt[, base::mean(x), by = grp]) %>% 
  data.frame %>% 
  select(bytes, calls)

#>      bytes                                                       calls
#> 1      280                          [() -> [.data.table() -> new.env()
#> 2     8240 [() -> [.data.table() -> null.data.table() -> setalloccol()
#> 3     8240 [() -> [.data.table() -> null.data.table() -> setalloccol()
#> 4  4000048                          [() -> [.data.table() -> forderv()
#> 5  2003000                                       [() -> [.data.table()
#> 6     8248                                       [() -> [.data.table()
#> 7  4005944                                       [() -> [.data.table()
#> 8     8248                                       [() -> [.data.table()
#> 9  2003000                                       [() -> [.data.table()
#> 10    8256                      [() -> [.data.table() -> setalloccol()
#> 11    8256                      [() -> [.data.table() -> setalloccol()

# benchmark without gforce
bench::mark(dt[, base::mean(x), by = grp], 
            iterations = 25) %>% 
  select(median, mem_alloc)

#> # A tibble: 1 x 2
#>     median mem_alloc
#>   <bch:tm> <bch:byt>
#> 1   10.5ms    11.5MB