Your example doesn’t work without allow.cartesian=TRUE.
You shouldn’t be using by=.EACHI here. This by was what was implicit in the earlier versions which made it slow. Please re-read the README.
Here’s the function I tested on 1.9.3:
calc1 <- function(d) {
d[ hit==1][ d,list(hittime=time),roll=-20, allow.cartesian=TRUE][ !is.na(hittime)]
}
calc2 <- function(d) {
temp <- d[ hit==1][ d,list(time),roll=-20, allow.cartesian=TRUE]
setnames(temp,1,"hittime")
temp[!is.na(hittime)]
}
# Generate sample data
set.seed(12312391)
data <- data.table(
group = sample(1e3,1e7,replace=T),
time = ceiling(runif(1e7, 0, 1e5)),
hit = rbinom(1e7, 1, p = 0.1),
key=c("group","time"))
system.time(ans1 <- calc1(data))
# user system elapsed
# 2.083 0.189 2.344
system.time(ans2 <- calc2(data))
# user system elapsed
# 2.012 0.241 2.426
identical(ans1, ans2) # [1] TRUE
You write:
I also don't see any way to refer to the different time vs. hittime without renaming the second time column.
I don’t quite follow what this means, but IIUC I think this is what you’re referring to: https://github.com/Rdatatable/data.table/issues/471
You write:
You mention some FR's, but they're hard to find without the specific numbers.
I was mentioning the first two points under NEW FEATURES within Changes in v1.9.3. The one that starts with by=.EACHI runs j for each group in x that each row of i joins to. and the one that starts with Accordingly, X[Y, j] now does what X[Y][, j] did.
Maybe we should start numbering the fixes for easy reference. Will note it down.
You write: Where can I find the 1.9.3 reference manual?
This version is a development version. Necesary changes will be reflected in their corresponding ?... entry. And when we find some time, the introduction and FAQs will be updated. But that’s not yet.
If you don’t wish to keep up-to-date by looking at the NEWS, you’ll have to wait until the next stable release on CRAN.
You write: On my system (MacOSX), build_vignettes=TRUE gives an error in texi2dvi -- would that have generated the refman? If so, how do I fix that?
I’m guessing it’s a PDF latex error. If so, you’ll have to install what the error message says is missing on your system. Sorry, can’t help you much there.
Arun
From: Stavros Macrakis (Σταῦρος Μακράκης) macrakis at alum.mit.edu
Reply: Stavros Macrakis (Σταῦρος Μακράκης) macrakis at alum.mit.edu
Date: June 30, 2014 at 10:40:24 PM
To: Arunkumar Srinivasan aragorn168b at gmail.com
Cc: datatable-help at r-forge.wu-wien.ac.at datatable-help at r-forge.wu-wien.ac.at
Subject: Re: [datatable-help] Speeding up column references with roll
OK, I'm retesting in 1.9.3, adding by=.EACHI. I don't see any significant difference in the timings -- setnames is still 25% faster than list(hittime=time). What exactly was fixed?
I also don't see any way to refer to the different time vs. hittime without renaming the second time column.
You mention some FR's, but they're hard to find without the specific numbers.
Where can I find the 1.9.3 reference manual? I think it would be easier to understand for me than the incremental changes in the New Features listings. On my system (MacOSX), build_vignettes=TRUE gives an error in texi2dvi -- would that have generated the refman? If so, how do I fix that?
Thanks,
-s
On Mon, Jun 30, 2014 at 1:00 PM, Arunkumar Srinivasan <aragorn168b at gmail.com> wrote:
Once again, has been fixed in 1.9.3. Now join requires `by=.EACHI` (explicit) to perform a by-without-by.
https://github.com/Rdatatable/data.table/blob/master/README.md
Have a look at the first FR (by = .EACHI runs ...) that's been fixed in 1.9.3 - there's some changes in the way join results in due to these changes (which've been discussed since and for quite sometime) to bring more consistency to the DT[i, j, by] syntax. Also have a look at the second FR and the links it points to for the discussions.
In general, it's better to test with the devel version (and have a look at README) for any bugs you may encounter.
Arun
From: Stavros Macrakis (Σταῦρος Μακράκης) macrakis at alum.mit.edu
Reply: Stavros Macrakis (Σταῦρος Μακράκης) macrakis at alum.mit.edu
Date: June 30, 2014 at 5:38:10 PM
To: datatable-help at r-forge.wu-wien.ac.at datatable-help at r-forge.wu-wien.ac.at
Subject: [datatable-help] Speeding up column references with roll
In the following example, it is about 15-25% faster to use setnames rather than j=list(name=var). Is there some better approach to referencing the other joined column when using roll?
# Use j=list(name=var)
calc1 <- function(d) {
d[ hit==1
][ d,list(hittime=time),roll=-20
][ !is.na(hittime)
]
}
# Use setnames
calc2 <- function(d) {
temp <- d[ hit==1
][ d,time,roll=-20
]
setnames(temp,3,"hittime")
temp[!is.na(hittime)]
}
# Generate sample data
set.seed(12312391)
data <- data.table(
group = sample(1e3,1e7,replace=T),
time = ceiling(runif(1e7, 0, 1e5)),
hit = rbinom(1e7, 1, p = 0.1),
key=c("group","time"))
# Timing
system.time(replicate(10,{gc();calc1(data)})) => 69 sec system.time(replicate(10,{gc();calc2(data)})) => 52 sec
_______________________________________________
datatable-help mailing list
datatable-help at lists.r-forge.r-project.org
https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/datatable-help
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/datatable-help/attachments/20140630/ae603c09/attachment.html>