arrow 3.0.0

Rust implementation of Apache Arrow

A native Rust implementation of Apache Arrow, a cross-language development platform for in-memory data.


Every Array in this crate has an associated DataType, that specifies how its data is layed in memory and represented. Thus, a central enum of this crate is DataType, that contains the set of valid DataTypes in the specification. For example, DataType::Utf8.


The central trait of this package is the dynamically-typed Array that represents a fixed-sized, immutable, Send + Sync Array of nullable elements. An example of such an array is UInt32Array. One way to think about an arrow Array is a Arc<[Option<T>; len]> where T can be anything ranging from an integer to a string, or even another Array.

Arrays have len(), data_type(), and the nullability of each of its elements, can be obtained via is_null(index). To downcast an Array to a specific implementation, you can use

use arrow::array::{Array, UInt32Array};
let array = UInt32Array::from(vec![Some(1), None, Some(3)]);
assert_eq!(array.len(), 3);
assert_eq!(array.value(0), 1);
assert_eq!(array.is_null(1), true);

To make the array dynamically typed, we wrap it in an Arc:

# use std::sync::Arc;
use arrow::datatypes::DataType;
use arrow::array::{UInt32Array, ArrayRef};
# let array = UInt32Array::from(vec![Some(1), None, Some(3)]);
let array: ArrayRef = Arc::new(array);
assert_eq!(array.len(), 3);
// array.value() is not available in the dynamically-typed version
assert_eq!(array.is_null(1), true);
assert_eq!(array.data_type(), &DataType::UInt32);

to downcast, use as_any():

# use std::sync::Arc;
# use arrow::array::{UInt32Array, ArrayRef};
# let array = UInt32Array::from(vec![Some(1), None, Some(3)]);
# let array: ArrayRef = Arc::new(array);
let array = array.as_any().downcast_ref::<UInt32Array>().unwrap();
assert_eq!(array.value(0), 1);

Memory and Buffers

Data in Array is stored in ArrayData, that in turn is a collection of other ArrayData and Buffers. Buffers is the central struct that array implementations use keep allocated memory and pointers. The MutableBuffer is the mutable counter-part ofBuffer. These are the lowest abstractions of this crate, and are used throughout the crate to efficiently allocate, write, read and deallocate memory.

Field, Schema and RecordBatch

Field is a struct that contains an array's metadata (datatype and whether its values can be null), and a name. Schema is a vector of fields with optional metadata. Together, they form the basis of a schematic representation of a group of Arrays.

In fact, RecordBatch is a struct with a Schema and a vector of Arrays, all with the same len. A record batch is the highest order struct that this crate currently offers and is broadly used to represent a table where each column in an Array.


This crate offers many operations (called kernels) to operate on Arrays, that you can find at [compute::kernels]. It has both vertical and horizontal operations, and some of them have an SIMD implementation.


This crate has most of the implementation of the arrow specification. Specifically, it supports the following types:

This crate also implements many common vertical operations:

as well as some horizontal operations, such as

Finally, this crate implements some readers and writers to different formats:

The parquet implementation is on a separate crate