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.

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.

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.

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.

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.

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

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.

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