GForce should be able to work with `:=` as well.

[arunsrinivasan]

No description provided.

[franknarf1]

Just ran into this today looking at a question on SO:

actions = data.table(User_id = c("Carl","Carl","Carl","Lisa","Moe"),
                     category = c(1,1,2,2,1),
                     value= c(10,20,30,40,50))
users = actions[, other_var := 1, by=User_id]

# verbose says: the following is not optimized
users[, value_one := 0 ]
users[actions[category==1], value_one := sum(value), on="User_id", by=.EACHI, verbose=TRUE]

# verbose says: the following is optimized
rbind( 
    actions[category==1], 
    unique(actions[,"User_id", with=FALSE])[, value := 0 ],
fill=TRUE)[, sum(value), by=User_id, verbose=TRUE]

To me, the first way looks idiomatic, considering the variable needs to end up in users in the end.

Another: https://stackoverflow.com/a/47338118/ (gtail)

Another https://stackoverflow.com/a/51569126/ should do DT[, mx := max(pt), by=Subject][, diff := mx - pt][] I guess

Another, specifically interested in memory performance: https://stackoverflow.com/q/52189712 "data.table reference semantics: memory usage of iterating through all columns"

Another, wants to scale/demean multiple variables: https://stackoverflow.com/q/52528123

Another taking max by group with a subsetting condition and adding with := (see akrun's answer) https://stackoverflow.com/a/54911855/ also related to the already-completed part of #971

[brodieG]

Just wanted to emphasize that enabling this can allow using GForce effectively for complex expressions, albeit with some work. For example I show in this post how to enable it for:

slope <- function(x, y) {
  x_ux <- x - mean(x)
  uy <- mean(y)
  sum(x_ux * (y - uy)) / sum(x_ux ^ 2)
}

By doing:

DT <- data.table(grp, x, y)
setkey(DT, grp)
DTsum <- DT[, .(ux=mean(x), uy=mean(y)), keyby=grp]
DT[DTsum, `:=`(x_ux=x - ux, y_uy=y - uy)]
DT[, `:=`(x_ux.y_uy=x_ux * y_uy, x_ux2=x_ux^2)]
DTsum <- DT[, .(x_ux.y_uy=sum(x_ux.y_uy), x_ux2=sum(x_ux2)), keyby=grp]
res.slope.dt2 <- DTsum[, .(grp, V1=x_ux.y_uy / x_ux2)]

Whereas if GForce was supported in := we could do:

DT <- data.table(grp, x, y)
DT[, `:=`(ux=mean(x), uy=mean(y)), keyby=grp]
DT[, `:=`(x_ux=x - ux, y_uy=y - uy)]
DT[, `:=`(x_ux.y_uy=x_ux * y_uy, x_ux2=x_ux^2)]
DTsum <- DT[, .(x_ux.y_uy=sum(x_ux.y_uy), x_ux2=sum(x_ux2)), keyby=grp]
res.slope.dt3 <- DTsum[, .(grp, x_ux.y_uy/x_ux2)]

Which looks cleaner and should be faster.

[brodieG]

Discussions with @MichaelChirico make me realize a very close cousin to this issue is:

>   DT <- data.table(x, y, grp)
>   DT[, .(x, mean(x)), keyby=grp]
Detected that j uses these columns: x 
Finding groups using forderv ... 1.049s elapsed (0.946s cpu) 
Finding group sizes from the positions (can be avoided to save RAM) ... 0.011s elapsed (0.011s cpu) 
lapply optimization is on, j unchanged as 'list(x, mean(x))'
GForce is on, left j unchanged
Old mean optimization changed j from 'list(x, mean(x))' to 'list(x, .External(Cfastmean, x, FALSE))'
Making each group and running j (GForce FALSE) ... 
  collecting discontiguous groups took 1.293s for 999953 groups
  eval(j) took 1.860s for 999953 calls
5.517s elapsed (3.862s cpu) 
              grp         x        V2
       1:       1 0.2151365 0.5512966
       2:       1 0.5358256 0.5512966
       3:       1 0.8496598 0.5512966
       4:       1 0.8480730 0.5512966
       5:       1 0.3464458 0.5512966
      ---                            
 9999996: 1000000 0.2601940 0.5474986
 9999997: 1000000 0.7940921 0.5474986
 9999998: 1000000 0.3825493 0.5474986
 9999999: 1000000 0.1786861 0.5474986
10000000: 1000000 0.9179119 0.5474986

Cross linking to #523.