37 releases (major breaking)

53.3.0 Nov 20, 2024
53.0.0 Sep 3, 2024
52.2.0 Jul 28, 2024
51.0.0 Mar 18, 2024
24.0.0 Oct 3, 2022

#90 in Data structures

Download history 321048/week @ 2024-08-19 312226/week @ 2024-08-26 283589/week @ 2024-09-02 311444/week @ 2024-09-09 275684/week @ 2024-09-16 313954/week @ 2024-09-23 304879/week @ 2024-09-30 343143/week @ 2024-10-07 307923/week @ 2024-10-14 368208/week @ 2024-10-21 367942/week @ 2024-10-28 311393/week @ 2024-11-04 305878/week @ 2024-11-11 286539/week @ 2024-11-18 194172/week @ 2024-11-25 218575/week @ 2024-12-02

1,015,212 downloads per month
Used in 465 crates (95 directly)

Apache-2.0

2MB
38K SLoC

The central type in Apache Arrow are arrays, which are a known-length sequence of values all having the same type. This crate provides concrete implementations of each type, as well as an Array trait that can be used for type-erasure.

Building an Array

Most Array implementations can be constructed directly from iterators or [Vec]

#
Int32Array::from(vec![1, 2]);
Int32Array::from(vec![Some(1), None]);
Int32Array::from_iter([1, 2, 3, 4]);
Int32Array::from_iter([Some(1), Some(2), None, Some(4)]);

StringArray::from(vec!["foo", "bar"]);
StringArray::from(vec![Some("foo"), None]);
StringArray::from_iter([Some("foo"), None]);
StringArray::from_iter_values(["foo", "bar"]);

ListArray::from_iter_primitive::<Int32Type, _, _>([
    Some(vec![Some(1), None, Some(3)]),
    None,
    Some(vec![])
]);

Additionally ArrayBuilder implementations can be used to construct arrays with a push-based interface

#
// Create a new builder with a capacity of 100
let mut builder = Int16Array::builder(100);

// Append a single primitive value
builder.append_value(1);
// Append a null value
builder.append_null();
// Append a slice of primitive values
builder.append_slice(&[2, 3, 4]);

// Build the array
let array = builder.finish();

assert_eq!(5, array.len());
assert_eq!(2, array.value(2));
assert_eq!(&array.values()[3..5], &[3, 4])

Low-level API

Internally, arrays consist of one or more shared memory regions backed by a Buffer, the number and meaning of which depend on the array’s data type, as documented in the Arrow specification.

For example, the type Int16Array represents an array of 16-bit integers and consists of:

Similarly, the type StringArray represents an array of UTF-8 strings and consists of:

  • An optional NullBuffer identifying any null values
  • An offsets OffsetBuffer<i32> identifying valid UTF-8 sequences within the values buffer
  • A values Buffer of UTF-8 encoded string data

Array constructors such as PrimitiveArray::try_new provide the ability to cheaply construct an array from these parts, with functions such as PrimitiveArray::into_parts providing the reverse operation.

#
// Create a Int32Array from Vec without copying
let array = Int32Array::new(vec![1, 2, 3].into(), None);
assert_eq!(array.values(), &[1, 2, 3]);
assert_eq!(array.null_count(), 0);

// Create a StringArray from parts
let offsets = OffsetBuffer::new(vec![0, 5, 10].into());
let array = StringArray::new(offsets, b"helloworld".into(), None);
let values: Vec<_> = array.iter().map(|x| x.unwrap()).collect();
assert_eq!(values, &["hello", "world"]);

As Buffer, and its derivatives, can be created from [Vec] without copying, this provides an efficient way to not only interoperate with other Rust code, but also implement kernels optimised for the arrow data layout - e.g. by handling buffers instead of values.

Zero-Copy Slicing

Given an Array of arbitrary length, it is possible to create an owned slice of this data. Internally this just increments some ref-counts, and so is incredibly cheap

let array = Int32Array::from_iter([1, 2, 3]);

// Slice with offset 1 and length 2
let sliced = array.slice(1, 2);
assert_eq!(sliced.values(), &[2, 3]);

Downcasting an Array

Arrays are often passed around as a dynamically typed &dyn Array or ArrayRef. For example, RecordBatch stores columns as ArrayRef.

Whilst these arrays can be passed directly to the compute, csv, json, etc... APIs, it is often the case that you wish to interact with the concrete arrays directly.

This requires downcasting to the concrete type of the array:


// Safely downcast an `Array` to an `Int32Array` and compute the sum
// using native i32 values
fn sum_int32(array: &dyn Array) -> i32 {
    let integers: &Int32Array = array.as_any().downcast_ref().unwrap();
    integers.iter().map(|val| val.unwrap_or_default()).sum()
}

// Safely downcasts the array to a `Float32Array` and returns a &[f32] view of the data
// Note: the values for positions corresponding to nulls will be arbitrary (but still valid f32)
fn as_f32_slice(array: &dyn Array) -> &[f32] {
    array.as_any().downcast_ref::<Float32Array>().unwrap().values()
}

The cast::AsArray extension trait can make this more ergonomic


fn as_f32_slice(array: &dyn Array) -> &[f32] {
    array.as_primitive::<Float32Type>().values()
}

Alternatives to ChunkedArray Support

The Rust implementation does not provide the ChunkedArray abstraction implemented by the Python and C++ Arrow implementations. The recommended alternative is to use one of the following:

  • Vec<ArrayRef> a simple, eager version of a ChunkedArray
  • impl Iterator<Item=ArrayRef> a lazy version of a ChunkedArray
  • impl Stream<Item=ArrayRef> a lazy async version of a ChunkedArray

Similar patterns can be applied at the RecordBatch level. For example, DataFusion makes extensive use of RecordBatchStream.

This approach integrates well into the Rust ecosystem, simplifies the implementation and encourages the use of performant lazy and async patterns.

use std::sync::Arc;
use arrow_array::{ArrayRef, Float32Array, RecordBatch, StringArray};
use arrow_array::cast::AsArray;
use arrow_array::types::Float32Type;
use arrow_schema::DataType;

let batches = [
   RecordBatch::try_from_iter(vec![
        ("label", Arc::new(StringArray::from(vec!["A", "B", "C"])) as ArrayRef),
        ("value", Arc::new(Float32Array::from(vec![0.1, 0.2, 0.3])) as ArrayRef),
    ]).unwrap(),
   RecordBatch::try_from_iter(vec![
        ("label", Arc::new(StringArray::from(vec!["D", "E"])) as ArrayRef),
        ("value", Arc::new(Float32Array::from(vec![0.4, 0.5])) as ArrayRef),
   ]).unwrap(),
];

let labels: Vec<&str> = batches
   .iter()
   .flat_map(|batch| batch.column(0).as_string::<i32>())
   .map(Option::unwrap)
   .collect();

let values: Vec<f32> = batches
   .iter()
   .flat_map(|batch| batch.column(1).as_primitive::<Float32Type>().values())
   .copied()
   .collect();

assert_eq!(labels, ["A", "B", "C", "D", "E"]);
assert_eq!(values, [0.1, 0.2, 0.3, 0.4, 0.5]);

Dependencies

~2.8–8.5MB
~71K SLoC