New class: data.table

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

New class: data.table

Matthew Dowle-4

Hi,

Following previous discussion on this list
(http://tolstoy.newcastle.edu.au/R/devel/05/12/3439.html) I have created a
package as suggested, and uploaded it to CRAN incoming : data.table.tar.gz.

** Your comments and feedback will be very much appreciated. **

>From help(data.table) :

This class really does very little. The only reason for its existence is
that the white book specifies that data.frame must have rownames.

Most of the code is copied from base functions with the code manipulating
row.names removed.

A data.table is identical to a data.frame other than:
        * it doesn't have rownames
        * [,drop] by default is FALSE, so selecting a single row will always
return a single row data.table not a vector
        * The comma is optional inside [], so DT[3] returns the 3rd row as a
1 row data.table
        * [] is like a call to subset()
        * [,...], is like a call to with().  (not yet implemented)

Motivation:
        * up to 10 times less memory
        * up to 10 times faster to create, and copy
        * simpler R code
        * the white book defines rownames, so data.frame can't be changed
... => new class

Examples:
nr = 1000000
D = rep(1:5,nr/5)
system.time(DF <<- data.frame(colA=D, colB=D))  # 2.08
system.time(DT <<- data.table(colA=D, colB=D))  # 0.15  (over 10 times
faster to create)
identical(as.data.table(DF), DT)
identical(dim(DT),dim(DF))
object.size(DF)/object.size(DT)                 # 10 times less memory

tt = subset(DF,colA>3)
ss = DT[colA>3]
identical(as.data.table(tt), ss)

mean(subset(DF,colA+colB>5,"colB"))
mean(DT[colA+colB>5]$colB)

tt = with(subset(DF,colA>3),colA+colB)
ss = with(DT[colA>3],colA+colB)                 # but could be:
DT[colA>3,colA+colB]  (not yet implemented)
identical(tt, ss)

tt = DF[with(DF,tapply(1:nrow(DF),colB,last)),] # select last row grouping
by colB
ss = DT[tapply(1:nrow(DT),colB,last)]           # but could be:
DT[last,group=colB]  (not yet implemented)
identical(as.data.table(tt), ss)

Lkp=1:3
tt = DF[with(DF,colA %in% Lkp),]              
ss = DT[colA %in% Lkp]                        # expressions inside the []
can see objects in the calling frame
identical(as.data.table(tt), ss)

In each case above there is either a space, time, or code brevity advantage
with the data.table.

The motivation for the new class grew from the realization that performance
of data.frames can be improved by removing the rownames.  See here for the
previous discussion
http://tolstoy.newcastle.edu.au/R/devel/05/12/3439.html.

Regards,
Matthew

______________________________________________
[hidden email] mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel