Marrow — Apache Arrow in Mojo

An implementation of Apache Arrow in Mojo. The initial motivation was to learn Mojo while doing something useful, and since I've been involved in Apache Arrow for a while it seemed a natural fit. The project has grown beyond a prototype: it now has a full Python binding layer, SIMD compute kernels, GPU acceleration, and benchmarks showing it outperforms PyArrow on array construction for common numeric and string workloads.

What is Arrow?

Apache Arrow is a cross-language development platform for in-memory data. It specifies a standardized, language-independent columnar memory format for flat and hierarchical data, organized for efficient analytic operations on modern hardware like CPUs and GPUs.

What is Mojo?

Mojo is a new programming language built on MLIR that combines Python expressiveness with the performance of systems programming languages.

Why Arrow in Mojo?

Arrow should be a first-class citizen in Mojo's ecosystem. This implementation provides zero-copy interoperability with PyArrow via the Arrow C Data Interface, and serves as a foundation for high-performance data processing in Mojo.

Features

Array types

• PrimitiveArray[T] — numeric and boolean arrays with type aliases: BoolArray, Int8Array … Int64Array, UInt8Array … UInt64Array, Float32Array, Float64Array
• StringArray — UTF-8 variable-length strings
• ListArray — variable-length nested arrays
• FixedSizeListArray — fixed-size nested arrays (embedding vectors, coordinates)
• StructArray — named-field structs
• ChunkedArray — array split across multiple chunks
• RecordBatch — schema + column arrays

Builders — incrementally build immutable arrays

• PrimitiveBuilder[T], StringBuilder, ListBuilder, FixedSizeListBuilder, StructBuilder
• AnyBuilder — type-erased builder using function-pointer vtable dispatch (O(1) copy via ArcPointer)

Compute kernels (SIMD-vectorized, null-aware)

• Arithmetic: add, sub, mul, div, neg, abs_, min_, max_
• Aggregates: sum_, product, min_, max_, any_, all_ (null-skipping)
• Selection: filter_, drop_nulls
• Similarity: cosine_similarity (batch N-vectors vs 1 query, CPU SIMD + GPU)

Python bindings — import marrow as ma

• array(values, type=None) — create any array type from Python lists with type inference
• All compute kernels exposed as free functions
• Full null handling, type coercion, nested structure support

Interoperability

• Arrow C Data Interface — zero-copy exchange with PyArrow
• GPU acceleration via Mojo's DeviceContext (Metal on Apple Silicon, CUDA on NVIDIA)

Python Quick Start

pixi run build_python   # compile marrow.so

import marrow as ma

# ── Array construction ────────────────────────────────────────────────────────

# Primitive arrays — type inference
a = ma.array([1, 2, 3, None, 5])           # int64 with one null
f = ma.array([1.0, 2.5, None, 4.0])        # float64

# Explicit types
a = ma.array([1, 2, 3, None, 5], type=ma.int64())

# Strings
s = ma.array(["hello", None, "world"])

# Nested lists
nested = ma.array([[1, 2], [3, 4, 5], None])

# Struct arrays — automatic type inference from dict keys
structs = ma.array([{"x": 1, "y": 1.5}, {"x": 2, "y": 2.5}])

# With explicit schema
t = ma.struct([ma.field("x", ma.int64()), ma.field("y", ma.float64())])
structs = ma.array([{"x": 1, "y": 1.5}, {"x": 2, "y": 2.5}], type=t)

# ── Arithmetic (null-propagating) ─────────────────────────────────────────────

b = ma.array([10, 20, 30, None, 50])
result = ma.add(a, b)      # null where either input is null
result = ma.sub(a, b)
result = ma.mul(a, b)
result = ma.div(a, b)

# ── Aggregates (null-skipping) ────────────────────────────────────────────────

ma.sum_(a)       # → 11.0  (skips the null at index 3)
ma.product(a)    # → 30.0
ma.min_(a)       # → 1.0
ma.max_(a)       # → 5.0
ma.any_(ma.array([False, True, None]))   # → True
ma.all_(ma.array([True, True, None]))   # → True

# ── Selection ─────────────────────────────────────────────────────────────────

mask = ma.array([True, False, True, False, True])
ma.filter_(a, mask)    # [1, 3, 5]
ma.drop_nulls(a)       # [1, 2, 3, 5]  (removes index 3)

# ── Array methods ─────────────────────────────────────────────────────────────

len(a)             # 5
a.null_count()     # 1
a.type()           # int64
a.slice(1, 3)      # [2, 3, None]  — zero-copy
a[0]               # 1
str(a)             # "Int64Array([1, 2, 3, NULL, 5])"

# Struct field access
structs.field(0)           # Int64Array — field "x"
structs.field("y")         # Float64Array — field "y"

Mojo API

Creating arrays

from marrow.arrays import array, PrimitiveArray, StringArray, BoolArray
from marrow.dtypes import int8, int32, int64, bool_, list_

# Factory function — list of optionals
var a = array[int32]([1, 2, 3, 4, 5])
var b = array[int64]([1, None, 3, None, 5])   # nulls at index 1 and 3
var c = array[bool_]([True, False, True])

Builders

from marrow.builders import PrimitiveBuilder, StringBuilder, ListBuilder

# Primitive
var pb = PrimitiveBuilder[int64](capacity=4)
pb.append(10)
pb.append(20)
pb.append_null()
pb.append(40)
var arr: Int64Array = pb.finish_typed()

# String
var sb = StringBuilder()
sb.append("hello")
sb.append_null()
sb.append("world")
var strs: StringArray = sb.finish_typed()

# List of int32 — append child elements, then commit each list element
var child = PrimitiveBuilder[int32]()
child.append(1)
child.append(2)
var lb = ListBuilder(child^)       # moves child into the builder
lb.append(True)                    # [1, 2] is the first list element
lb.values().append(3)              # child element for the next list
lb.append(True)                    # [3] is the second list element
lb.append_null()                   # null third element
var lists: ListArray = lb.finish_typed()

Display

All arrays implement Writable so they print directly:

print(arr)    # Int64Array([10, 20, NULL, 40])
print(strs)   # StringArray([hello, NULL, world])

Compute kernels

from marrow.kernels.arithmetic import add, sub, mul, div
from marrow.kernels.aggregate import sum_, min_, max_, any_, all_
from marrow.kernels.filter import filter_, drop_nulls

var x = array[int64]([1, 2, 3, 4])
var y = array[int64]([10, 20, 30, 40])

var z = add(x, y)               # Int64Array([11, 22, 33, 44])
var total = sum_[int64](x)      # 10
var filtered = filter_[int64](x, array[bool_]([True, False, True, False]))

Zero-copy PyArrow interop (C Data Interface)

from python import Python
from marrow.c_data import CArrowArray, CArrowSchema

var pa = Python.import_module("pyarrow")
var pyarr = pa.array([1, 2, 3, 4, 5], mask=[False, False, False, False, True])

var c_array = CArrowArray.from_pyarrow(pyarr)
var c_schema = CArrowSchema.from_pyarrow(pyarr.type)

var dtype = c_schema.to_dtype()     # int64
var data = c_array.to_array(dtype)
var typed = data.as_int64()

print(typed.is_valid(0))   # True
print(typed.is_valid(4))   # False  (null)
print(typed.unsafe_get(0)) # 1

Benchmarks

Python array construction vs PyArrow (n=100,000 elements, Apple M-series, mean time):

Array type	marrow	PyArrow	speedup
int64 (explicit type)	0.37 ms	0.89 ms	2.4x faster
int64 + nulls (explicit)	0.36 ms	0.87 ms	2.4x faster
float64 (explicit)	0.34 ms	0.48 ms	1.4x faster
float64 + nulls	0.34 ms	0.51 ms	1.5x faster
string (explicit)	0.72 ms	1.06 ms	1.5x faster
string + nulls	0.70 ms	1.04 ms	1.5x faster
struct, primitive fields	5.41 ms	6.54 ms	1.2x faster
int64 (inferred)	1.40 ms	1.27 ms	1.1x slower
string (inferred)	1.57 ms	1.02 ms	1.5x slower
nested list (inferred)	3.92 ms	2.36 ms	1.7x slower

When the array type is provided explicitly, marrow's builder path is faster than PyArrow's for numeric and string types. Type inference involves a Python-side scan to detect the type, which adds overhead; this gap will narrow as the inference path is optimized.

Run the benchmarks yourself:

pixi run bench_python       # Python array construction vs PyArrow
pixi run bench              # CPU SIMD arithmetic benchmarks
pixi run bench_similarity   # cosine similarity: CPU vs GPU

GPU Acceleration

GPU kernels are available for compute-intensive operations when a DeviceContext is provided. Benchmarked on Apple Silicon (M-series, Metal, unified memory):

Cosine similarity (batch N-vectors vs 1 query, dim=768):

Vectors	CPU SIMD	GPU (upload per call)	GPU (pre-loaded)
10 K	baseline	2–3x slower	~1x (crossover)
100 K	baseline	~1x	~3x faster
500 K	baseline	—	~13x faster

The key pattern: upload data to the GPU once, run multiple kernels, download results at the end. The crossover vs CPU SIMD is around 10K vectors at dim≥384.

Element-wise arithmetic (add, mul, etc.) is faster on CPU SIMD — data transfer overhead dominates for low arithmetic-intensity operations.

from std.gpu.host import DeviceContext
from marrow.kernels.similarity import cosine_similarity

# Pre-load data onto the GPU once
var ctx = DeviceContext()
var vectors_gpu = vectors.to_device(ctx)
var query_gpu = query.to_device(ctx)

# Run many similarity searches without re-uploading
var scores = cosine_similarity(vectors_gpu, query_gpu, ctx)

Known Limitations

1. C Data Interface: Release callbacks are not invoked (Mojo cannot pass a callback to a C function yet). Consuming Arrow data from PyArrow works; producing data back to PyArrow via the release mechanism is not fully implemented.

2. Testing: Conformance against the Arrow specification is verified through PyArrow since Mojo has no JSON library yet. Full integration testing requires a Mojo JSON reader.

3. Type coverage: Only boolean, numeric, string, list, fixed-size list, and struct types are implemented. Date/time, dictionary, union, decimal, and binary types are not yet supported.

4. GPU null handling: Binary arithmetic kernels on the GPU do not propagate null bitmaps (GPU bitmap_and is not yet implemented). Null-aware GPU arithmetic is CPU-only for now.

Development

Install pixi, then:

pixi run test              # run all tests (Mojo + Python)
pixi run test_mojo         # Mojo unit tests only
pixi run test_python       # Python binding tests only
pixi run bench             # CPU/GPU arithmetic benchmarks
pixi run bench_python      # Python vs PyArrow array construction benchmarks
pixi run bench_similarity  # cosine similarity: CPU vs GPU vs GPU preloaded
pixi run fmt               # format all code (Mojo + Python)

If the project matures, the goal is to contribute it upstream to the Apache Arrow project.

References

• Arrow columnar format specification
• Arrow C Data Interface
• Another effort to implement Arrow in Mojo