shift() in data.table v1.9.6 is slow for many groups

[nachti]

Hi there!
For many different groups in by, shift is much slower than manual shifting.
See: http://stackoverflow.com/questions/35179911/shift-in-data-table-v1-9-6-is-slow-for-many-groups
and https://github.com/nachti/datatable_test/blob/master/leadtest.R for a detailed example.
Cheers,
Gerhard

[arunsrinivasan]

That's not surprising. This'll go away when gforce is optimised for :=. It's on the list for this release, I believe.

[ben519]

+1 for this performance enhancement. shift() is the main bottleneck in a lot of my code. Seems that for a fixed number of rows, the time it takes to run shift() is proportional to the number of groups in the data.

library(data.table)

# Build table to store timings
timings <- CJ(RowCount = 10^7, Groups = 10^c(0:7))
timings[, SizePerGroup := RowCount/Groups]

# Loop through each experiment
for(i in 1:nrow(dt)){
  print(paste0("Iteration: ", i))
  
  # Build dataset
  timings_i <- timings[i]
  dt <- data.table(Grp = rep(seq_len(timings_i$Groups), each = timings_i$SizePerGroup))
  dt[, Value := sample(100, size = .N, replace = T)]
  
  # Measure the time it takes to insert a column indicating the previous value by group
  elapsed <- system.time(dt[, PrevValueByGrp := shift(Value, type = "lag"), by = Grp])["elapsed"]
  timings[i, Elapsed := elapsed]
}

library(ggplot2)
ggplot(timings, aes(x = Groups, y = Elapsed))+geom_line()+geom_point()

[franknarf1]

@ben519 Fyi, for the special case of when your code looks like that, there's a shortcut:

library(data.table)
dt <- data.table(Grp = rep(seq_len(1e6), each=10L))
dt[, Value := sample(100L, size = .N, replace = TRUE)]

system.time(dt[, PrevValueByGrp := shift(Value, type = "lag"), by = Grp][])
#    user  system elapsed 
#   19.50    0.80   20.34
system.time(dt[, v := shift(Value, type = "lag")][rowid(Grp)==1L, v := NA][])
#    user  system elapsed 
#    1.00    0.87    1.25 

dt[, all.equal(v, PrevValueByGrp)]
# [1] TRUE