11 releases (5 breaking)

new 0.6.2 Oct 9, 2021
0.6.1 Oct 7, 2021
0.5.3 Sep 14, 2021
0.4.0 Aug 24, 2021
0.1.0 Jul 20, 2021

#2 in #analytics

Download history 49/week @ 2021-07-16 113/week @ 2021-07-23 48/week @ 2021-07-30 24/week @ 2021-08-06 19/week @ 2021-08-13 149/week @ 2021-08-20 287/week @ 2021-08-27 65/week @ 2021-09-03 797/week @ 2021-09-10 1172/week @ 2021-09-17 658/week @ 2021-09-24 1013/week @ 2021-10-01 1001/week @ 2021-10-08

1,524 downloads per month
Used in 14 crates (4 directly)


45K SLoC

Arrow2: Transmute-free Arrow

test codecov

This repository contains a Rust library to work with the Arrow format. It is a re-write of the official Arrow crate using transmute-free operations. See FAQ for details.

See the guide for a general introduction to this crate and its general components, and API docs for a detailed documentation of each of its APIs.


This repo and crate's primary goal is to offer a safe Rust implementation of the Arrow specification. As such, it

  • MUST NOT implement any logical type other than the ones defined on the arrow specification, schema.fbs.
  • MUST lay out memory according to the arrow specification
  • MUST support reading from and writing to the C data interface at zero-copy.
  • MUST support reading from, and writing to, the IPC specification, which it MUST verify against golden files available here.

Design documents about each of the parts of this repo are available on their respective READMEs.

Check DEVELOPMENT.md for our development practices.


The test suite is a superset of all tests that the original implementation has against golden files from the arrow project. It includes both little and big endian files.

Furthermore, the CI runs all integration tests against apache/arrow@master, demonstrating full interoperability with other implementations.

Finally, we have integration tests against parquet generated by pyarrow under different configurations, as well as integration tests against pyspark demonstrating compatibility with its parquet reader.


We use the SemVer 2.0 used by Cargo and the remaining of the Rust ecosystem, we also use the 0.x.y versioning, since we are iterating over the API.

Features in this crate and not in the official

Safety and Security

  • safe by design (i.e. no transmutes, runtime type checking nor pointer casts)
  • Uses Rust's compiler whenever possible to prove that memory reads are sound
  • All non-IO components pass MIRI checks (MIRI can't open files atm)

Arrow Format

  • IPC supports big endian
  • MutableArray API to work in-memory in-place.
  • faster IPC reader (different design that avoids an extra copy of all data)
  • IPC supports 2.0 (compression)
  • Extension type supported
  • All implemented arrow types pass FFI integration tests against pyarrow / C++
  • All implemented arrow types pass IPC integration tests against other implementations


  • Reading parquet is 10-20x faster (single core) and deserialization is parallelizable
  • Writing parquet is 3-10x faster (single core) and serialization is parallelizable
  • parquet IO has no unsafe
  • parquet IO supports async


  • More predictable JSON reader
  • Generalized parsing of CSV based on logical data types

Features in the original not available in this crate

  • Parquet read and write of struct and nested lists.

Features in this crate not in pyarrow

  • Read and write of delta-encoded utf8 to and from parquet
  • parquet roundtrip of all supported arrow types.

Features in pyarrow not in this crate

Too many to enumerate; e.g. nested dictionary arrays, nested parquet.



The arrow crate uses Buffer, a generic struct to store contiguous memory regions (of bytes). This construct is used to store data from all arrays in the rust implementation. The simplest example is a buffer containing 1i32, that is represented as &[0u8, 0u8, 0u8, 1u8] or &[1u8, 0u8, 0u8, 0u8] depending on endianness.

When a user wishes to read from a buffer, e.g. to perform a mathematical operation with its values, it needs to interpret the buffer in the target type. Because Buffer is a contiguous region of bytes with no type information, users must transmute its data into the respective type.

Arrow currently transmutes buffers on almost all operations, and very often does not verify that there is type alignment nor correct length when we transmute it to a slice of type &[T].

Just as an example, in v5.0.0, the following code compiles, does not panic, is unsound and results in UB:

let buffer = Buffer::from_slic_ref(&[0i32, 2i32])
let data = ArrayData::new(DataType::Int64, 10, 0, None, 0, vec![buffer], vec![]);
let array = Float64Array::from(Arc::new(data));

println!("{:?}", array.value(1));

Note how this initializes a buffer with bytes from i32, initializes an ArrayData with dynamic type Int64, and then an array Float64Array from Arc<ArrayData>. Float64Array's internals will essentially consume the pointer from the buffer, re-interpret it as f64, and offset it by 1.

Still within this example, if we were to use ArrayData's datatype, Int64, to transmute the buffer, we would be creating &[i64] out of a buffer created out of i32.

Any Rust developer acknowledges that this behavior goes very much against Rust's core premise that a functions' behvavior must not be undefined depending on whether the arguments are correct. The obvious observation is that transmute is one of the most unsafe Rust operations and not allowing the compiler to verify the necessary invariants is a large burden for users and developers to take.

This simple example indicates a broader problem with the current design, that we now explore in detail.

Root cause analysis

At its core, Arrow's current design is centered around two main structs:

  1. untyped Buffer
  2. untyped ArrayData
1. untyped Buffer

The crate's buffers are untyped, which implies that once created, the type information is lost. Consequently, the compiler has no way of verifying that a certain read can be performed. As such, any read from it requires an alignment and size check at runtime. This is not only detrimental to performance, but also very cumbersome.

For the past 4 months, I have identified and fixed more than 10 instances of unsound code derived from the misuse, within the crate itself, of Buffer. This hints that downstream dependencies using this crate and use this API are likely do be even more affected by this.

2. untyped ArrayData

ArrayData is a struct containing buffers and child data that does not differentiate which type of array it represents at compile time.

Consequently, all buffer reads from ArrayData's buffers are effectively unsafe, as they require certain invariants to hold. These invariants are strictly related to ArrayData::datatype: this enum differentiates how to transmute the ArrayData::buffers. For example, an ArrayData::datatype equal to DataType::UInt32 implies that the buffer should be transmuted to u32.

The challenge with the above struct is that it is not possible to prove that ArrayData's creation is sound at compile time. As the sample above showed, there was nothing wrong, during compilation, with passing a buffer with i32 to an ArrayData expecting i64. We could of course check it at runtime, and we should, but we are defeating the whole purpose of using a typed system as powerful as Rust offers.

The main consequence of this observation is that the current code has a significant maintenance cost, as we have to be rigorously check the types of the buffers we are working with. The example above shows that, even with that rigour, we fail to identify obvious problems at runtime.

Overall, there are many instances of our code where we expose public APIs marked as safe that are unsafe and lead to undefined behavior if used incorrectly. This goes against the core goals of the Rust language, and significantly weakens Arrow Rust's implementation core premise that the compiler and borrow checker proves many of the memory safety concerns that we may have.

Equally important, the inability of the compiler to prove certain invariants is detrimental to performance. As an example, the implementation of the take kernel in this repo is semantically equivalent to the current master, but 1.3-2x faster.


Contrarily to the original implementation, this implementation does not transmutate byte buffers based on runtime types, and instead requires all buffers to be typed (in Rust's sense of a generic).

This removes many existing bugs and enables the compiler to prove all type invariants with the only exception of FFI and IPC boundaries.

This crate also has a different design towards arrays' offsets that removes many out of bound reads consequent of using byte and slot offsets interchangibly.

This crate's design of primitive types is also more explicit about its logical and physical representation, enabling support for Timestamps with timezones and a safe implementation of the Interval type.

Consequently, this crate is easier to use, develop, maintain, and debug.

Any plans to merge with the Apache Arrow project?

Maybe. The primary reason to have this repo and crate is to be able to propotype and mature using a fundamentally different design based on a transmute-free implementation. This requires breaking backward compatibility and loss of features that is impossible to achieve on the Arrow repo.

Furthermore, the arrow project currently has a release mechanism that is unsuitable for this type of work:

  • A release of the Apache consists of a release of all implementations of the arrow format at once, with the same version. It is currently at 5.0.0.

This implies that the crate version is independent of the changelog or its API stability, which violates SemVer. This procedure makes the crate incompatible with Rusts' (and many others') ecosystem that heavily relies on SemVer to constraint software versions.

Secondly, this implies the arrow crate is versioned as >0.x. This places expectations about API stability that are incompatible with this effort.


Licensed under either of

at your option.


Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.


~90K SLoC