The Apache Arrow C++ library provides rich, powerful features for working with columnar data. The
arrow R package provides both a low-level interface to the C++ library and some higher-level, R-flavored tools for working with it. This vignette provides an overview of how the pieces fit together, and it describes the conventions that the classes and methods follow in R.
arrow package lets you work efficiently with large, multi-file datasets
dplyr methods. See
vignette("dataset", package = "arrow") for an overview.
arrow provides some simple functions for using the Arrow C++ library to read and write files.
These functions are designed to drop into your normal R workflow
without requiring any knowledge of the Arrow C++ library
and use naming conventions and arguments that follow popular R packages, particularly
The readers return
(or if you use the
tibble package, they will act like
and the writers take
arrow provides basic read and write support for the Apache
Parquet columnar data file format.
library(arrow) df <- read_parquet("path/to/file.parquet")
Just as you can read, you can write Parquet files:
arrow package also includes a faster and more robust implementation of the
Feather file format, providing
write_feather(). This implementation depends
on the same underlying C++ library as the Python version does,
resulting in more reliable and consistent behavior across the two languages, as
well as improved performance.
arrow also by default writes the Feather V2 format,
which supports a wider range of data types, as well as compression.
For CSV and line-delimited JSON, there are
read_csv_arrow() currently has fewer parsing options for dealing with
every CSV format variation in the wild, for the files it can read, it is
often significantly faster than other R CSV readers, such as
arrow lets you
share data between R and Python (
pyarrow) efficiently, enabling you to take
advantage of the vibrant ecosystem of Python packages that build on top of
Apache Arrow. See
vignette("python", package = "arrow") for details.
arrow package also provides many lower-level bindings to the C++ library, which enable you
to access and manipulate Arrow objects. You can use these to build connectors
to other applications and services that use Arrow. One example is Spark: the
sparklyr package has support for using Arrow to
move data to and from Spark, yielding significant performance
Arrow defines the following classes for representing metadata:
|Class||Description||How to create an instance|
||attribute controlling how values are represented||functions in
||a character string name and a
Arrow defines the following classes for representing zero-dimensional (scalar), one-dimensional (array/vector-like), and two-dimensional (tabular/data frame-like) data:
|Dim||Class||Description||How to create an instance|
||single value and its
||vector of values and its
||vectors of values and their
Each of these is defined as an
R6 class in the
arrow R package and
corresponds to a class of the same name in the Arrow C++ library. The
package provides a variety of
R6 and S3 methods for interacting with instances
of these classes.
For convenience, the
arrow package also defines several synthetic classes that
do not exist in the C++ library, including:
ArrowDatum: inherited by
ArrowTabular: inherited by
ArrowObject: inherited by all Arrow objects
Arrow has a rich data type system that includes direct parallels with R's data types and much more.
In the tables, entries with a
- are not currently implemented.
|R type||Arrow type|
1: If the character vector exceeds 2GB of strings, it will be converted to a
large_utf8 Arrow type
2: Only lists where all elements are the same type are able to be translated to Arrow list type (which is a “list of” some type).
|Arrow type||R type|
3: These integer types may contain values that exceed the range of R's
integer type (32-bit signed integer). When they do,
uint64 are converted to
double (“numeric”) and
int64 is converted to
bit64::integer64. This conversion can be disabled (so that
int64 always yields a
bit64::integer64 vector) by setting
options(arrow.int64_downcast = FALSE).
4: Due to the limitation of R
dictionary values are coerced to string when translated to R if they are not already strings.
arrow*_binary classes are implemented as lists of raw vectors.
arrow*_list classes are implemented as subclasses of
vctrs_list_of with a
ptype attribute set to what an empty Array of the value type converts to.
Arrow supports custom key-value metadata attached to Schemas. When we convert a
data.frame to an Arrow Table or RecordBatch, the package stores any
attributes() attached to the columns of the
data.frame in the Arrow object's Schema. These attributes are stored under the “r” key; you can assign additional string metadata under any other key you wish, like
x$metadata$new_key <- "new value".
This metadata is preserved when writing the table to Feather or Parquet, and when reading those files into R, or when calling
as.data.frame() on a Table/RecordBatch, the column attributes are restored to the columns of the resulting
data.frame. This means that custom data types, including
vctrs annotations, and others, are preserved when doing a round-trip through Arrow.
Note that the
attributes() stored in
$metadata$r are only understood by R. If you write a
haven columns to a Feather file and read that in Pandas, the
haven metadata won't be recognized there. (Similarly, Pandas writes its own custom metadata, which the R package does not consume.) You are free, however, to define custom metadata conventions for your application and assign any (string) values you want to other metadata keys. For more details, see the documentation for
C++ is an object-oriented language, so the core logic of the Arrow library is encapsulated in classes and methods. In the R package, these classes are implemented as
R6 reference classes, most of which are exported from the namespace.
In order to match the C++ naming conventions, the
R6 classes are in TitleCase, e.g.
RecordBatch. This makes it easy to look up the relevant C++ implementations in the code or documentation. To simplify things in R, the C++ library namespaces are generally dropped or flattened; that is, where the C++ library has
arrow::io::FileOutputStream, it is just
FileOutputStream in the R package. One exception is for the file readers, where the namespace is necessary to disambiguate. So
Some of these classes are not meant to be instantiated directly; they may be base classes or other kinds of helpers. For those that you should be able to create, use the
$create() method to instantiate an object. For example,
rb <- RecordBatch$create(int = 1:10, dbl = as.numeric(1:10)) will create a
RecordBatch. Many of these factory methods that an R user might most often encounter also have a
snake_case alias, in order to be more familiar for contemporary R users. So
record_batch(int = 1:10, dbl = as.numeric(1:10)) would do the same as
The typical user of the
arrow R package may never deal directly with the
R6 objects. We provide more R-friendly wrapper functions as a higher-level interface to the C++ library. An R user can call
read_parquet() without knowing or caring that they're instantiating a
ParquetFileReader object and calling the
$ReadFile() method on it. The classes are there and available to the advanced programmer who wants fine-grained control over how the C++ library is used.