vctrs

Travis build status Coverage status lifecycle

There are three main goals to the vctrs package, each described in a vignette:

vctrs is a developer focused package. Understanding and extending vctrs requires some effort from developers, but should be invisible to most users. It’s our hope that having an underlying theory will mean that users can build up an accurate mental model without explicitly learning the theory. vctrs will typically be used by other packages, making it easy for them to provide new classes of S3 vectors that are supported throughout the tidyverse (and beyond). For that reason, vctrs has few dependencies.

Installation

vctrs is not currently on CRAN. Install the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("r-lib/vctrs")

Usage

library(vctrs)

# Prototypes
str(vec_type_common(FALSE, 1L, 2.5))
#>  num(0)
str(vec_cast_common(FALSE, 1L, 2.5))
#> List of 3
#>  $ : num 0
#>  $ : num 1
#>  $ : num 2.5

# Sizes
str(vec_size_common(1, 1:10))
#>  int 10
str(vec_recycle_common(1, 1:10))
#> List of 2
#>  $ : num [1:10] 1 1 1 1 1 1 1 1 1 1
#>  $ : int [1:10] 1 2 3 4 5 6 7 8 9 10

Motivation

The original motivation for vctrs from two separate, but related problems. The first problem is that base::c() has rather undesirable behaviour when you mix different S3 vectors:

# combining factors makes integers
c(factor("a"), factor("b"))
#> [1] 1 1

# combing dates and date-times give incorrect values
dt <- as.Date("2020-01-1")
dttm <- as.POSIXct(dt)

c(dt, dttm)
#> [1] "2020-01-01"    "4321940-06-07"
c(dttm, dt)
#> [1] "2019-12-31 18:00:00 CST" "1969-12-31 23:04:22 CST"

This behaviour arises because c() has dual purposes: as well as it’s primary duty of combining vectors, it has a secondary duty of stripping attributes. For example, ?POSIXct suggests that you should use c() if you want to reset the timezone.

The second problem is that dplyr::bind_rows() is not extensible by others. Currently, it handles arbitrary S3 classes using heuristics, but these often fail, and it feels like we really need to think through the problem in order to build a principled solution. This intersects with the need to cleanly support more types of data frame columns including lists of data frames, data frames, and matrices.