Block Records and Row Records

  collapse = TRUE,
  comment = "#>"

Block Records and Row Records

The theory of cdata data transforms is based on the principles:

The idea of data coordinates is related to Codd's 2nd rule:

Each and every datum (atomic value) in a relational data base is guaranteed to be logically accessible by resorting to a combination of table name, primary key value and column name.

The coordinatized data concept is that the exact current data realization is incidental. One can perform a data change of basis to get the data into the right format (where the physical layout of records is altered to match the desired logical layout of the data).

The idea of data records (and these records possibly being different than simple rows) is a staple of computer science: harking at least back to record-oriented filesystems.

The core of the cdata package is to supply transforms between what we call "row records" (records that happen to be implemented as a single row) and block records (records that span multiple rows). These two methods are:

All the other cdata functions are helpers allowing abbreviated notation in special cases (such as unpivot_to_blocks() pivot_to_rowrecs()) and adapters (allowing these operations to be performed directly in databases and large data systems such as Apache Spark).

The current favored idiomatic interfaces to cdata are:

Let's look at cdata with some specific data.

For our example let's take the task of re-organizing the iris data for a faceted plot, as discussed here.


iris <- data.frame(iris)
iris$iris_id <- seq_len(nrow(iris))

head(iris, n=1)

To transform this data into a format ready for our ggplot2 task we design (as taught here) a "transform control table" that shows how to move from our row-oriented form into a block oriented form. Which in this case looks like the following.

In R the transform table is specified as follows.

controlTable <- wrapr::qchar_frame(
  "flower_part", "Length"      , "Width"     |
  "Petal"      , Petal.Length  , Petal.Width |
  "Sepal"      , Sepal.Length  , Sepal.Width )

layout <- rowrecs_to_blocks_spec(
  recordKeys = c("iris_id", "Species"))


And then applying it converts rows from our iris data into ready to plot 2-row blocks.

iris %.>%
  head(., n = 1) %.>%

iris_aug <- iris %.>%
  head(., n = 1) %.>%

iris_aug %.>%

To perform the reverse transform we use the same transform control table, but we think of it as specifying the reverse transform (from its own block form into a row). The reverse can be specified using the t() transpose/adjoint method.

# re-do the forward transform, this time
# with more records so we can see more
iris_aug <- iris %.>%
  head(., n = 3) %.>%


inv_layout <- t(layout)


# demonstrate the reverse transform
iris_back <- iris_aug %.>%


cdata considers the row-record a universal intermediate form, and this has the advantage of being able to represent a different type per value (as each value per-record is in a different column)

This differs from reshape2 where the melt() to "molten" (or thin RDF-triple-like) is used as the universal intermediate form that one then dcast()s into desired arrangements.

As we have said, a tutorial on how to design a controlTable can be found here and here.

Some additional (older) tutorials on cdata data transforms can are given below:


The cdata operators can be related to Codd's relational operators as follows:

Try the cdata package in your browser

Any scripts or data that you put into this service are public.

cdata documentation built on Aug. 20, 2023, 1:06 a.m.