Rerun repeated uniqueN test

[mattdowle]

Was reported here : #3395 (comment)
I said I'd follow up here: #3435 (comment)
Double check all those results under the new default.

[jangorecki]

Up to date timings below. There is still a problem in uniqueN by group.
We basically need what @mattdowle described in #3743 (comment) and #1120 would also be nice.

busy machine (20 cores)

#[1] 10
#   user  system elapsed 
#172.084   1.640  18.179 

#[1] 1
#   user  system elapsed 
# 10.608   0.596  11.206 

#[1] 20
#still computing

idle machine (32 cores)

#[1] 16
#   user  system elapsed 
#258.708   0.876  16.638  

#[1] 1
#   user  system elapsed 
#  8.405   0.132   8.539 

#[1] 32
#    user   system  elapsed 
#1066.565    1.180   35.085 

code

library(data.table)
N_X = 1e6
n_day = 60
n_clientid = 1e5
n_Platform = 7
X = data.table(
  day = sample(1:n_day, N_X, TRUE),
  clientid = as.character(sample(1:n_clientid, N_X, TRUE)),
  Platform = as.character(sample(1:n_Platform, N_X, TRUE))
)

setDTthreads(NULL) # default
getDTthreads()
system.time(
  X[, .(x = uniqueN(day) - 1L,
        first_active_day = min(day),
        last_active_day = max(day)),
    by = .(Platform, clientid)]
)

setDTthreads(1)
getDTthreads()
system.time(
  X[, .(x = uniqueN(day) - 1L,
        first_active_day = min(day),
        last_active_day = max(day)),
    by = .(Platform, clientid)]
)

setDTthreads(0)
getDTthreads()
system.time(
  X[, .(x = uniqueN(day) - 1L,
        first_active_day = min(day),
        last_active_day = max(day)),
    by = .(Platform, clientid)]
)

[mattdowle]

Closed by #4484. See benchmark near the top of the top comment here: #4484 (comment)