Fast, minimal, and ergonomic Apache Arrow implementation in Rust, with PyO3 and Python bindings.

Highlights:

• < 2s compilation times (0.15s rebuilds)
• Rust to Python for 1m rows in under 220ns¹
• Guaranteed 64-byte SIMD alignment
• Fully-typed throughout
• Embed Python ML in Rust
• Plug in Polars, Arrow, and Python with Zero-copy FFI, PyCapsule and PYO3
• NdArray and XArray implementations with native DLPack integration for bridging to Python AI/ML workloads.

Limitations:

• No Arrow lists or structs.

Why Minarrow?

• Productive: Lean base dependencies help keep flow and library fast.
• Fully typed: Enum dispatch ensures constant feedback through a strong compiler + IDE feedback loop
• Fail fast: Errors that might slip through a dynamic dispatch boundary into run-time errors, stay compile-time.
• Ergonomics: Convenient dot syntax, composable opt-up abstractions, and flexible signatures via liberal From trait implementations.
• Speed: builds on the Vec64 crate so data is fully SIMD compatible for an extra layer of low-latency parallelism, with out of the box numeric, bitmask and string SIMD kernels to help get you started.
• Minimal Dependencies: Small security surface - built from foundations. Only log, num-traits (and custom-built Vec64) in the base build.

Installation

cargo add minarrow

Minarrow uses the nightly toolchain for allocator_api and portable_simd:

rustup override set nightly

Quick Start

use minarrow::{arr_i32, arr_f64, arr_str32, arr_bool, MaskedArray};

// Create arrays with macros
let ids = arr_i32![1, 2, 3, 4];
let prices = arr_f64![10.5, 20.0, 15.75];
let names = arr_str32!["alice", "bob", "charlie"];
let flags = arr_bool![true, false, true];

// Direct typed access - no downcasting
assert_eq!(ids.len(), 4);
assert_eq!(prices.num().f64().get(0), Some(10.5));

use minarrow::{tbl, fa_i32, fa_str32, Print};

// Build tables via FieldArrays with constructor macros
let table = tbl!(
    "users",
    fa_i32!("id", 1, 2, 3),
    fa_str32!("name", "alice", "bob", "charlie"),
);
table.print();

Core Features

Typed Arrays

Six typed arrays back standard workloads:

Semantic groupings (NumericArray, TextArray, TemporalArray) support flexibility call-site signatures and static dispatch.

Array and Table (RecordBatch) sit on top, with chunked Super (Vec<Table>) versions for streaming.

Bonus LAPACK-compatible Matrix and Cube types support analytical workload variations.

NdArray n-dimensional numeric data, with chunked (SuperNdArray) and labelled (XArray) forms, and hands tensors to PyTorch, JAX, and NumPy zero-copy over DLPack.

Zero-Copy Views

Ergonomic zero-copy row and column selection.

use minarrow::*;

let table = create_table();

// Pandas-style selection
let view = table.c(&["name", "value"]);    // columns
let view = table.r(10..20);                // rows
let view = table.c(&["A", "B"]).r(0..100); // both

// Materialise when needed
let owned = view.to_table();

Off-the-Wire pattern

use std::sync::Arc;
use minarrow::{st, Consolidate, SuperTable};

// Append batches as they arrive
let mut stream = SuperTable::new("trades".into());
stream.push(Arc::new(batch1));
stream.push(Arc::new(batch2));

// Assemble from existing batches
let stream = st!("trades", batch1, batch2);

// Consolidate to a single table when ready
let table = stream.consolidate();

Arrow Interop

Embed Python in Rust. Bench in /examples.

// Run a Random Forest Classifier using Python in Rust
let value = rt.with_python(&dataset, |py, obj| {
    let scope = PyDict::new(py);
    scope.set_item("table", obj)?;
    py.run(
        cr#"
        import polars as pl
        from sklearn.ensemble import RandomForestClassifier
        from sklearn.model_selection import train_test_split

        df = table.to_polars()
        features = ["x0", "x1", "x2", "x3"]
        X = df.select(features).to_numpy()
        y = df["label"].to_numpy()

        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=0)
        model = RandomForestClassifier(n_estimators=100, random_state=0)
        model.fit(X_train, y_train)
        predicted = model.predict(X_test)

        result = pl.DataFrame(
            {
                "actual": y_test.astype("int64"),
                "predicted": predicted.astype("int64"),
            }
        )
        "#,
        Some(&scope),
        Some(&scope),
    )?;
    scope
        .get_item("result")?
        .ok_or_else(|| pyo3::exceptions::PyKeyError::new_err("result not set"))
})?;

Normal Rust - use Polars, Arrow, convert at the boundary, stay native internally:

// To Arrow (feature: cast_arrow)
let batch = table.to_apache_arrow();          // RecordBatch
let column = table.cols[0].to_apache_arrow(); // ArrayRef

// To Polars (feature: cast_polars)
let df = table.to_polars();                   // DataFrame

// FFI via the Arrow C Data Interface
use minarrow::ffi::arrow_c_ffi::export_to_c;
let (array_ptr, schema_ptr) = export_to_c(array, schema);

Architecture

Minarrow uses enums for type dispatch instead of trait objects.

// Static dispatch routing, full compiler inlining
match array {
    Array::NumericArray(num) => match num {
        NumericArray::Int64(arr) => process(arr),
        NumericArray::Float64(arr) => process(arr),
        // ...
    },
    // ...
}

// Get fully typed IntegerArray<i64> instead of Array:
let array = array.num().i64();

Benchmarks

Array performance

Sum of 1,000 integers, averaged over 1,000 runs¹:

Minarrow's direct access is within the noise threshold of raw Vec performance whilst maintaining SIMD-compatible alignment.

Python Roundtrip

• GIL acquisition (uncontended): 53ns
• Rust to Python¹ 1m rows, 2 columns : 165ns
• Python to Rust¹ 1m rows, 2 columns: 2.5μs

See minarrow-pyo3/examples

Test machine

▎ ¹ = Intel Core Ultra 7 155H · 32 GB · Ubuntu 24.04 · 1.97-nightly release build.

Feature Flags

Default features: views, chunked, large_string, simd, select, log.

Interop:

Additional types:

Performance:

Extras:

See Cargo.toml for the full list with detailed notes on each.

Ecosystem

Contributing

Contributions are welcome, particularly in the following areas:

1. Nested types - List and Struct support
2. Bug fixes

All contributions are subject to the Contributor Licence Agreement (CLA). See CONTRIBUTING.md for details.

License

Copyright © 2025–2026 Peter Garfield Bower.

Released under the Apache 2.0 License. See LICENSE for details.

Acknowledgements

Minarrow is a from-scratch implementation of the Apache Arrow memory layout inspired by the standards pioneered by Apache Arrow, Arrow2, and Polars.

Minarrow is not affiliated with Apache Arrow.

SpaceCell

Minarrow is maintained by SpaceCell and forms part of its open-source foundation for high-performance data computing.