Introduction

This vignette motivates why we wrote git2rdata and illustrates how you can use it to store dataframes as plain text files.

Maintaining Variable Classes

R has different options to store dataframes as plain text files from R. Base R has write.table() and its companions like write.csv(). Some other options are data.table::fwrite(), readr::write_delim(), readr::write_csv() and readr::write_tsv(). Each of them writes a dataframe as a plain text file by converting all variables into characters. After reading the file, they revert this conversion. The distinction between character and factor gets lost in translation. read.table() converts by default all strings to factors, readr::read_csv() keeps by default all strings as character. These functions cannot recover the factor levels. These functions determine factor levels based on the observed levels in the plain text file. Hence factor levels without observations will disappear. The order of the factor levels is also determined by the available levels in the plain text file, which can be different from the original order.

The write_vc() and read_vc() functions from git2rdata keep track of the class of each variable and, in case of a factor, also of the factor levels and their order. Hence this function pair preserves the information content of the dataframe. The vc suffix stands for version control as these functions use their full capacity in combination with a version control system.

Efficiency Relative to Storage and Time

Optimizing File Storage

Plain text files require more disk space than binary files. This is the price we have to pay for a readable file format. The default option of write_vc() is to create file as compact as possible. Since we use a tab delimited file format, we can omit quotes around character variables. This saves 2 bytes per row for each character variable. write_vc add quotes automatically in the exceptional cases when we needed them, e.g. to store a string that contains tab or newline characters. We don’t add quotes to row-variable combinations where we don’t need them.

Since we store the class of each variable, we can further reduce the file size by following rules:

  • Store a logical as 0 (FALSE), 1 (TRUE) or NA to the data.
  • Store a factor as its indices in the data. Store the index, labels of levels and their order in the metadata.
  • Store a POSIXct as a numeric to the data. Store the class and the origin in the metadata. Store and return timestamps as UTC.
  • Store a Date as an integer to the data. Store the class and the origin in the metadata.

Storing the factors, POSIXct and Date as their index, makes them less user readable. The user can turn off this optimization when user readability is more important than file size.

Optimized for Version Control

Another main goal of git2rdata is to optimise the storage of the plain text files under version control. write_vc() and read_vc() has methods for interacting with git repositories using the git2r framework. Users who want to use git without git2r or use a different version control system (e.g. Subversion, Mercurial), still can use git2rdata to write the files to disk and uses their preferred workflow on version control.

Hence, write_vc() will always perform checks to look for changes which potentially lead to large diffs. More details on this in vignette("version_control", package = "git2rdata"). Some problems will always yield a warning. Other problems will yield an error by default. The user can turn these errors into warnings by setting the strict = FALSE argument.

As this vignette ignores the part on version control, we will always use write_vc(strict = FALSE) and hide the warnings to improve the readability.

Basic Usage

Let’s start by setting up the environment. We need a directory to store the data and a dataframe to store.

# Create a directory in tempdir
path <- tempfile(pattern = "git2r-")
dir.create(path)
# Create dummy data
set.seed(20190222)
x <- data.frame(
  x = sample(LETTERS),
  y = factor(
    sample(c("a", "b", NA), 26, replace = TRUE),
    levels = c("a", "b", "c")
  ),
  z = c(NA, 1:25),
  abc = c(rnorm(25), NA),
  def = sample(c(TRUE, FALSE, NA), 26, replace = TRUE),
  timestamp = seq(
    as.POSIXct("2018-01-01"),
    as.POSIXct("2019-01-01"),
    length = 26
  ),
  stringsAsFactors = FALSE
)
str(x)
#> 'data.frame':    26 obs. of  6 variables:
#>  $ x        : chr  "V" "U" "Z" "W" ...
#>  $ y        : Factor w/ 3 levels "a","b","c": 1 2 NA NA 1 NA 2 1 NA 1 ...
#>  $ z        : int  NA 1 2 3 4 5 6 7 8 9 ...
#>  $ abc      : num  -0.382 -0.42 -0.917 0.387 -0.992 ...
#>  $ def      : logi  TRUE FALSE NA FALSE NA NA ...
#>  $ timestamp: POSIXct, format: "2018-01-01 00:00:00" "2018-01-15 14:24:00" ...

Storing Optimized

Use write_vc() to store the dataframe. The root argument refers to the base directory where we store the data. The file argument becomes the base name of the files. The data file gets a .tsv extension, the metadata file a .yml extension. file can include a relative path starting from root.

library(git2rdata)
write_vc(x = x, file = "first_test", root = path, strict = FALSE)
#> 7fed986e0f5c218e3d0a4f0cc434bd87bb1e5f82 f8350dc218051af4bafcd8872d92b1a29cbb4f31 
#>                         "first_test.tsv"                         "first_test.yml"

write_vc() returns a vector of relative paths to the raw data and metadata files. The names of this vector contains the hashes of these files. We can have a look at both files. We’ll display the first 10 rows of the raw data. Notice that the YAML format of the metadata has the benefit of being both human and machine readable.

print_file <- function(file, root, n = -1) {
  fn <- file.path(root, file)
  data <- readLines(fn, n = n)
  cat(data, sep = "\n")
}
print_file("first_test.tsv", path, 10)
#> x    y   z   abc def timestamp
#> V    1   NA  -0.382010380419258  1   1514764800
#> U    2   1   -0.420347607856041  0   1516026240
#> Z    NA  2   -0.916731402237418  NA  1517287680
#> W    NA  3   0.387455128525654   0   1518549120
#> L    1   4   -0.992354993526956  NA  1519810560
#> C    NA  5   0.0228713954429028  NA  1521072000
#> R    2   6   -0.947557467717088  1   1522333440
#> S    1   7   -0.16302914628615   NA  1523594880
#> O    NA  8   0.523643352634392   1   1524856320
print_file("first_test.yml", path)
#> ..generic:
#>   git2rdata: 0.2.1
#>   optimize: yes
#>   NA string: NA
#>   hash: f8350dc218051af4bafcd8872d92b1a29cbb4f31
#>   data_hash: 7fed986e0f5c218e3d0a4f0cc434bd87bb1e5f82
#> x:
#>   class: character
#> 'y':
#>   class: factor
#>   labels:
#>   - a
#>   - b
#>   - c
#>   index:
#>   - 1
#>   - 2
#>   - 3
#>   ordered: no
#> z:
#>   class: integer
#> abc:
#>   class: numeric
#> def:
#>   class: logical
#> timestamp:
#>   class: POSIXct
#>   origin: 1970-01-01 00:00:00
#>   timezone: UTC

Storing Verbose

Adding optimize = FALSE to write_vc() will keep the raw data in a human readable format. The metadata file is slightly different. The most obvious is the optimize: no tag and the different hash. Another difference is the metadata for POSIXct and Date classes. They will no longer have an origin tag but a format tag.

write_vc(x = x, file = "verbose", root = path, optimize = FALSE, strict = FALSE)
#> d1a61b91465221b40f7986c2a8da421a2e166c5b 9fc40476f0ba0c4cb4225d5313d56760a6d3b065 
#>                            "verbose.tsv"                            "verbose.yml"
print_file("verbose.tsv", path, 10)
#> x    y   z   abc def timestamp
#> V    a   NA  -0.382010380419258  TRUE    2018-01-01T00:00:00Z
#> U    b   1   -0.420347607856041  FALSE   2018-01-15T14:24:00Z
#> Z    NA  2   -0.916731402237418  NA  2018-01-30T04:48:00Z
#> W    NA  3   0.387455128525654   FALSE   2018-02-13T19:12:00Z
#> L    a   4   -0.992354993526956  NA  2018-02-28T09:36:00Z
#> C    NA  5   0.0228713954429028  NA  2018-03-15T00:00:00Z
#> R    b   6   -0.947557467717088  TRUE    2018-03-29T14:24:00Z
#> S    a   7   -0.16302914628615   NA  2018-04-13T04:48:00Z
#> O    NA  8   0.523643352634392   TRUE    2018-04-27T19:12:00Z
print_file("verbose.yml", path)
#> ..generic:
#>   git2rdata: 0.2.1
#>   optimize: no
#>   NA string: NA
#>   hash: 9fc40476f0ba0c4cb4225d5313d56760a6d3b065
#>   data_hash: d1a61b91465221b40f7986c2a8da421a2e166c5b
#> x:
#>   class: character
#> 'y':
#>   class: factor
#>   labels:
#>   - a
#>   - b
#>   - c
#>   index:
#>   - 1
#>   - 2
#>   - 3
#>   ordered: no
#> z:
#>   class: integer
#> abc:
#>   class: numeric
#> def:
#>   class: logical
#> timestamp:
#>   class: POSIXct
#>   format: '%Y-%m-%dT%H:%M:%SZ'
#>   timezone: UTC

Efficiency Relative to File Storage

Storing dataframes optimized or verbose has an impact on the required file size. The efficiency vignette give a comparison.

Reading Data

You retrieve the data with read_vc(). This function will reinstate the variables to their original state.

y <- read_vc(file = "first_test", root = path)
all.equal(x, y, check.attributes = FALSE)
#> [1] TRUE
y2 <- read_vc(file = "verbose", root = path)
all.equal(x, y2, check.attributes = FALSE)
#> [1] TRUE

read_vc() requires the meta data. It cannot handle dataframe not stored by write_vc().

Missing Values

write_vc() has an na argument which specifies the string which to use for missing values. Because we avoid using quotes, this string must be different from any character value in the data. This includes factor labels with verbose data storage. write_vc() checks this and will always return an error, even with strict = FALSE.

write_vc(x, "custom_na", path, strict = FALSE, na = "X", optimize = FALSE)
#> Error: one of the strings matches the NA string ('X')
#> Please use a different NA string or consider using a factor.
write_vc(x, "custom_na", path, strict = FALSE, na = "b", optimize = FALSE)
#> Error: one of the levels matches the NA string ('b').
#> Please use a different NA string or use optimize = TRUE
write_vc(x, "custom_na", path, strict = FALSE, na = "X")
#> Error: one of the strings matches the NA string ('X')
#> Please use a different NA string or consider using a factor.
write_vc(x, "custom_na", path, strict = FALSE, na = "b")
#> 5415c0176d1c25fc1ac08763ae9271f6306eccc5 3c028ddbb9efabc8c85a396b2db1f4f1c8d9461d 
#>                          "custom_na.tsv"                          "custom_na.yml"

Please note that write_vc() uses the same NA string for the entire dataset, thus for every variable.

print_file("custom_na.tsv", path, 10)
#> x    y   z   abc def timestamp
#> V    1   b   -0.382010380419258  1   1514764800
#> U    2   1   -0.420347607856041  0   1516026240
#> Z    b   2   -0.916731402237418  b   1517287680
#> W    b   3   0.387455128525654   0   1518549120
#> L    1   4   -0.992354993526956  b   1519810560
#> C    b   5   0.0228713954429028  b   1521072000
#> R    2   6   -0.947557467717088  1   1522333440
#> S    1   7   -0.16302914628615   b   1523594880
#> O    b   8   0.523643352634392   1   1524856320
print_file("custom_na.yml", path, 4)
#> ..generic:
#>   git2rdata: 0.2.1
#>   optimize: yes
#>   NA string: b

The default string for missing values is "NA". We recommend to keep this default, as long as the dataset permits it. A first good alternative is an empty string (""). If that won’t work either, you’ll have to use your imagination. Try to keep it short, clear and robust1.

write_vc(x, "custom_na", path, strict = FALSE, na = "")
#> dfaf9f13491b30a88b7e37dabd9f9c87225fbba0 d2b145b74dbe6bd1d7883a62f2549c124988ec63 
#>                          "custom_na.tsv"                          "custom_na.yml"

  1. robust in the sense that you won’t need to change it later