Upcoming versions of base R eliminate the need for DT() functionality - consider eliminating?

[msummersgill]

As an experimental feature the placeholder _ can now also be used in the ‘rhs’ of a forward pipe |> expression as the first argument in an extraction call, such as _$coef. More generally, it can be used as the head of a chain of extractions, such as _$coef[[2]].

R-devel/News, Mon, 13 Feb 2023

In upcoming versions of R, syntax such as the following will be supported:

as.data.table(mtcars) |> 
  _[am == 1] |> 
  _[, .(maxhp = max(hp)), by = .(cyl)]

#    cyl maxhp
# 1:   6   175
# 2:   4   113
# 3:   8   335

With this in mind, I think it's worth reconsidering whether DT() should remain on the release roadmap.

Some arguments in favor of scrapping DT() are as follows:

• Base R functionality dt |> _[...] will address all use cases for DT()
• Introducing an alternate syntax for calling [.data.table will be confusing for new users. Advertising this functionality also eliminates an opportunity to educate users about existing dt[...][...] chaining capabilities baked in by default
• Exporting DT() will lead to namespace collision issues with the DT::DT() exported by the widely used DT package
• DT() acceptance of non data.table objects is likely to cause user confusion
  • Unclear class for argument and return value of DT #5559
  • The new function DT() used for chaining/piping drops data.frame rownames names #5430
  • Naming conflict and unexpected behavior with the functional form of data.table DT() #5129
• Supporting DT() going forward will be a drain on valuable maintainer's time - DT() Labeled Issues
• DT() has not yet been exported in a CRAN release so no there will no impacts to reverse dependencies and limited impact to users who have adopted this form

That being said, this is just my two cents as a satisfied user of data.table interested in the long term success of the package. Happy to hear counter arguments, and open to the idea that the broader community may see enough value to finish the push to support DT().

[tdhock]

I agree it would be good to remove if possible, to avoid user confusion, and to save dev time.

[MichaelChirico]

given the number of headaches induced by what was supposed to be a "simple" wrapper, I would also be very happy to drop DT().

that said, I would pause on this issue until that syntax lands in a released version.

[MichaelChirico]

@markfairbanks care to elaborate? it will be important feedback.

[markfairbanks]

I would love to have a DT() that would work on older versions of R (whether to v4.0.0 with the base pipe or with v3.4.0 with the magrittr pipe). It's good that they're adding the _ placeholder for the base pipe but that won't help very much with backwards compatibility for users that don't always have access to the newest version of R.

And as far as the current issues - every issue linked in the original comment goes back to the fact that DT() as currently constructed tries to work on data.frames. It seems much more logical to have DT() only work on data.tables instead of removing it entirely. Having it expand to data.frame capability can be added in a later version if people really want it (and then there would be more time to work out the kinks). Or if it's not something that's fixable it can just work on data.tables into the future.

Something like this:

library(data.table)

DT <- function(x, ...) {
  if (!is.data.table(x)) {
    stop("`x` must be a data.table")
  }
  x[...]
}

df <- data.table(x = 1:3, y = c("a", "a", "b"))

copy(df) |>
  DT(, double_x := x * 2) |>
  DT()
#>        x      y double_x
#>    <int> <char>    <num>
#> 1:     1      a        2
#> 2:     2      a        4
#> 3:     3      b        6

There wouldn't be anything "surprising" with how it works because it would work exactly like [.data.table but work better with piping.

[msummersgill]

that said, I would pause on this issue until that syntax lands in a released version.
@MichaelChirico

Agree - but perhaps might be a reason to hold back DT() from CRAN releases in the interim? It looks like this change made it into R 4.3 (news item is included in changelog of release candidate R-latest.tar.gz), which has a planned release date of April 21st, 2023, so there shouldn't be too long of a wait.

I would love to have a DT() that would work on older versions of R (whether to v4.0.0 with the base pipe or with v3.4.0 with the magrittr pipe).
@markfairbanks

Since backporting |> isn't an option, the only portable solution is to use magrittr, which has always allowed [ operations on the . placeholder. Does that address the desire here?

## Works R > 3.4 using magrittr
as.data.table(mtcars) %>%
  .[am == 1] %>% 
  .[, .(maxhp = max(hp)), by = .(cyl)]

[markfairbanks]

I think it still makes more sense to have a functional DT() - a version that is designed to work with pipes without any extra thought or placeholders 🤷‍♂️ In general the placeholder seems like a workaround, not integrated functionality.

DT() also makes everything much more approachable for newer R users trying to use piping with data.table or dplyr users making the switch.

[eddelbuettel]

As we all got used to working with the dev version of data.table() I too had some use cases for DT() and grew reasonably fond of it -- in particular for one (only partially done) project where I translated some code 'from that other paradigm'.

So with the caveat that I am not fully up to speed on any implementation headaches, my $0.02 are for shipping DT() as it is a nice usability improvement, especially for newbs.

[Kamgang-B]

I think that it would be useful to export DT() only when it will be mature enough to properly handle data.frame objects in general (DT() tested enough and all pending issues solved). Given the number of related issues, exporting DT() should not necessarily be considered for the upcoming data.table releases as it will take an important amount of time to solve all of them.

I personally believe that the most important benefit of DT() is the fact that it can be used with data.frame objects in general and not only with data.tables. If only data.table object is considered, I think that the upcoming R version eliminates the need for DT().

But I believe that DT() has one benefit that is not eliminated by the upcoming R version: it allows the use of data.table syntax on data.frame objects in general. I think that one of the main limitations of data.table so far is related to the fact that its syntax works only with data.table objects. So, solving the issues related to DT() will remove this limitation.

While piping using the square brackets (repeatedly) was hard to follow and kept indenting the codes to the right, the upcoming R version literally solves this issue. I really like the fact that the square brackets are vertically aligned (with no additional indentation).

In conclusion, I would suggest to keep DT() only if it will also support data.frames in general (and not only data.tables); otherwise, I think that the upcoming R eliminates the need for DT().

[jangorecki]

Whenever size is not a problem you can just alias DT=as.data.table and use [

[tdhock]

another couple pieces of information that may be relevant to this discussion

• There was at some point in the past another/different DT, https://cloud.r-project.org/web/packages/data.table/vignettes/datatable-faq.html#DTremove1
• DT=data.table(some.example.data) is used frequently in documentation materials, so the new (old?) DT function would conflict with that (potentially confusing).

[tdhock]

if we do keep DT, can someone (@markfairbanks or?) please volunteer to maintain it and review changes to it, by adding your name to the new/proposed CODEOWNERS file? #5629 (currently DT is defined in data.table.R but maybe would be better to move it to a separate file, to make it clear who is responsible for it)

[jangorecki]

Anyway DT() is not going to land on CRAN as long as pending issues related to it are not being addressed. Putting it on CRAN having those issue would be a nightmare, and most likely new issues would be detected as well.
For new coming (non-patch) release what is currently planned is to un-export DT() as described in #5472