The goal of narrow is to wrap the Arrow Data C API and Arrow Stream C API to provide lightweight Arrow support for R packages to consume and produce streams of data in Arrow format.
Installation
You can install the development version from GitHub with:
# install.packages("remotes")
remotes::install_github("paleolimbot/narrow")
Creating arrays
You can create an Arrow array using as_narrow_array()
. For many types (e.g., integers and doubles), this is done without any copying of memory: narrow just arranges the existing R vector memory and protects it for the lifetime of the underlying struct ArrowArray
.
library(narrow)
(array <- as_narrow_array(1:5))
#> <narrow_array i[5]>
#> - schema:
#> <narrow_schema 'i' at 0x55967f8285a0>
#> - format: i
#> - name:
#> - flags: nullable
#> - metadata: list()
#> - dictionary: NULL
#> - children[0]:
#> - array_data:
#> <narrow_array_data at 0x559682692480>
#> - length: 5
#> - null_count: 0
#> - offset: 0
#> - buffers[2]: List of 2
#> $ : NULL
#> $ : int [1:5] 1 2 3 4 5
#> - dictionary: NULL
#> - children[0]:
For Array
s and RecordBatch
es from the arrow package, this is almost always a zero-copy operation and is instantaneous even for very large Arrays.
library(arrow)
(array2 <- as_narrow_array(Array$create(1:5)))
#> <narrow_array i[5]>
#> - schema:
#> <narrow_schema 'i' at 0x5596817a4350>
#> - format: i
#> - name:
#> - flags: nullable
#> - metadata: list()
#> - dictionary: NULL
#> - children[0]:
#> - array_data:
#> <narrow_array_data at 0x559681bbade0>
#> - length: 5
#> - null_count: 0
#> - offset: 0
#> - buffers[2]: List of 2
#> $ :<externalptr>
#> $ :<externalptr>
#> - dictionary: NULL
#> - children[0]:
Exporting arrays
To convert an array object to some other type, use from_narrow_array()
:
str(from_narrow_array(array))
#> int [1:5] 1 2 3 4 5
The narrow package has built-in defaults for converting arrays to R objects; you can also specify your own using the ptype
argument:
str(from_narrow_array(array, ptype = double()))
#> num [1:5] 1 2 3 4 5
from_narrow_array(array, ptype = arrow::Array)
#> Array
#> <int32>
#> [
#> 1,
#> 2,
#> 3,
#> 4,
#> 5
#> ]
Streams
The Arrow C API also specifies an experimental stream interface. In addition to handling streams created elsewhere, you can create streams based on a narrow_array()
:
stream1 <- as_narrow_array_stream(as_narrow_array(1:3))
narrow_array_stream_get_next(stream1)
#> <narrow_array i[3]>
#> - schema:
#> <narrow_schema 'i' at 0x559682878380>
#> - format: i
#> - name:
#> - flags: nullable
#> - metadata: list()
#> - dictionary: NULL
#> - children[0]:
#> - array_data:
#> <narrow_array_data at 0x5596812cff20>
#> - length: 3
#> - null_count: 0
#> - offset: 0
#> - buffers[2]: List of 2
#> $ :<externalptr>
#> $ :<externalptr>
#> - dictionary: NULL
#> - children[0]:
narrow_array_stream_get_next(stream1)
#> NULL
…or based on a function that returns one or more narrow_array()
s:
counter <- -1
rows_per_chunk <- 5
csv_file <- readr::readr_example("mtcars.csv")
schema <- as_narrow_array(
readr::read_csv(
csv_file,
n_max = 0,
col_types = readr::cols(.default = readr::col_double())
)
)$schema
stream2 <- narrow_array_stream_function(schema, function() {
counter <<- counter + 1L
result <- readr::read_csv(
csv_file,
skip = 1 + (counter * rows_per_chunk),
n_max = rows_per_chunk,
col_names = c(
"mpg", "cyl", "disp", "hp", "drat",
"wt", "qsec", "vs", "am", "gear", "carb"
),
col_types = readr::cols(.default = readr::col_double())
)
if (nrow(result) > 0) result else NULL
})
You can pass these to Arrow as a RecordBatchReader
using narrow_array_stream_to_arrow()
:
reader <- narrow_array_stream_to_arrow(stream2)
as.data.frame(reader$read_table())
#> # A tibble: 32 × 11
#> mpg cyl disp hp drat wt qsec vs am gear carb
#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 21 6 160 110 3.9 2.62 16.5 0 1 4 4
#> 2 21 6 160 110 3.9 2.88 17.0 0 1 4 4
#> 3 22.8 4 108 93 3.85 2.32 18.6 1 1 4 1
#> 4 21.4 6 258 110 3.08 3.22 19.4 1 0 3 1
#> 5 18.7 8 360 175 3.15 3.44 17.0 0 0 3 2
#> 6 18.1 6 225 105 2.76 3.46 20.2 1 0 3 1
#> 7 14.3 8 360 245 3.21 3.57 15.8 0 0 3 4
#> 8 24.4 4 147. 62 3.69 3.19 20 1 0 4 2
#> 9 22.8 4 141. 95 3.92 3.15 22.9 1 0 4 2
#> 10 19.2 6 168. 123 3.92 3.44 18.3 1 0 4 4
#> # … with 22 more rows
Round-turn operations for RecordBatch
also work:
df <- readr::read_csv(csv_file, show_col_types=FALSE)
as.data.frame(from_narrow_array(as_narrow_array(df), arrow::RecordBatch))
#> # A tibble: 32 × 11
#> mpg cyl disp hp drat wt qsec vs am gear carb
#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 21 6 160 110 3.9 2.62 16.5 0 1 4 4
#> 2 21 6 160 110 3.9 2.88 17.0 0 1 4 4
#> 3 22.8 4 108 93 3.85 2.32 18.6 1 1 4 1
#> 4 21.4 6 258 110 3.08 3.22 19.4 1 0 3 1
#> 5 18.7 8 360 175 3.15 3.44 17.0 0 0 3 2
#> 6 18.1 6 225 105 2.76 3.46 20.2 1 0 3 1
#> 7 14.3 8 360 245 3.21 3.57 15.8 0 0 3 4
#> 8 24.4 4 147. 62 3.69 3.19 20 1 0 4 2
#> 9 22.8 4 141. 95 3.92 3.15 22.9 1 0 4 2
#> 10 19.2 6 168. 123 3.92 3.44 18.3 1 0 4 4
#> # … with 22 more rows
C data access
The C data interface is ABI stable (and a version of the stream interface will be ABI stable in the future) so you can access the underlying pointers in compiled code from any R package (or inline C or C++ code). A narrow_schema()
is an external pointer to a struct ArrowSchema
, a narrow_array_data()
is an external pointer to a struct ArrowArray
, and a narrow_array()
is a list()
of a narrow_schema()
and a narrow_array_data()
.
#include <R.h>
#include <Rinternals.h>
#include "narrow.h"
SEXP extract_null_count(SEXP array_data_xptr) {
struct ArrowArray* array_data = (struct ArrowArray*) R_ExternalPtrAddr(array_data_xptr);
return Rf_ScalarInteger(array_data->null_count);
}
.Call("extract_null_count", as_narrow_array(c(NA, NA, 1:5))$array_data)
#> [1] 2
The lifecycles of objects pointed to by the external pointers are managed by R’s garbage collector: any object that gets garbage collected has its release()
callback called (if it isn’t NULL
) and the underlying memory for the struct Arrow...
freed. You can call the release()
callback yourself from compiled code but you probably don’t want to unless you’re explicitly limiting access to an object.